id stringlengths 36 36 | source stringclasses 15 values | formatted_source stringclasses 13 values | text stringlengths 2 7.55M |
|---|---|---|---|
94ac0691-c74f-4730-a28b-7e2a7d0a70e3 | trentmkelly/LessWrong-43k | LessWrong | Parameter vs Synapse?
In terms of AI timelines, the biggest question I haven't seen addressed is the computational equivalence of a synapse vs a parameter in modern neural nets. This seems like a very important input for any prediction on when we will have human-level AI.
Moravec's estimates of a retina vs the then-current edge-detection methods are sort of worthless under the assumption that AI will be built using modern learning methods, because feature engineered code is plausibly much more compute efficient than learned policies on tasks that are comprehensible to human programmers, and Moravec compared the retina with the former.
To pump this intuition, if we assume that 1 synapse == 1 parameter, then Moravec's estimates are more than 6 orders of magnitude too low. The size of the largest models we are able to train, on computers that are more powerful than Moravec predicted are needed for AGI, is at most 10 billion parameters, which is about as many synapses as the retina has. Though, it would be more accurate to say we are memory bandwidth constrained, as we train neural networks at much higher temporal resolution than human brains.
A very interesting question, then, is how many parameters does a modern learned model need to have plausibly similar capabilities to a human retina. This seems hard to investigate but not impossible. If we look at the first few layers of a conv net, for example, they seem to be doing the sort of edge detection that the retina does.
I think this would be a high-leverage thing to investigate as both the parameter > synapse and parameter < synapse are not implausible at first glance. Some people think the brain is doing something mysterious and better than backpropagation. Some people think it is just doing a worse, Rube Goldberg approximation of backpropagation. Artificial spiking neural networks perform really poorly compared standard stuff, which may imply paramater > synapse, etc.
If we assume AGI will come from scaling up current methods (rather |
cd01d7a4-c090-4a9b-97a8-111709a4c479 | trentmkelly/LessWrong-43k | LessWrong | Will Values and Competition Decouple?
There are a great many forces shaping the evolution of the universe. Among them, the values of agents -- systems which attempt to optimize, or steer the future towards certain configurations over others -- seem likely to have a dominant influence on the long-term future. The values of the agents around now have been largely determined by competitive pressures. Many people in the rationalist/EA community seem to take it for granted that this is soon going to change, and we will enter an era in which values and competition are completely decoupled; the values of the beings around at the time of this decoupling will be "locked in" and determine the shape of the entire future. I think is it plausible(>30% probability) that they are wrong, and that competition will continue, with at least some strength, indefinitely. If this is true, it has major implications for the likely trajectory of the world and how we should go about influencing the long-term future. In this blog post I'll lay out why I think this and what the implications are.
Epistemic status: not confident that the thesis is correct; I am confident that the community should be allocating more probability mass to this scenario than they currently are. If you like, imagine prepending every statement with "there is at least a 30% probability that".
SUMMARY
1. I sketch three possible scenarios for what the value systems of machine intelligences might look like. In two of these scenarios, values and competition are totally decoupled; in the third, they remain partially coupled.
2. I present the most basic arguments for and against the occurrence of decoupling. Briefly, the difficulty of ensuring successor alignment might generate competitive pressure towards value systems that try to accrue power to their successors in a value-agnostic way. I define autopoietic agents, systems which increase the number and influence of systems similar to themself.
3. I survey some more arguments given in the EA/rationalist c |
05c665e7-696f-4f45-b56b-c340a413bbec | trentmkelly/LessWrong-43k | LessWrong | The WASH-1400 Reactor Safety Study
In Intelligence Explosion analysis draft: introduction, Luke Muehlhauser and Anna Salamon wrote
> Norman Rasmussen's (1975) analysis of the safety of nuclear power plants, written before any nuclear accidents had occurred, correctly predicted several details of the Three Mile Island incident that previous experts had not (McGrayne 2011, 180).
I investigated this further a part of the project "Can we know what to do about AI?". The analysis that Muehlhauser and Salamon reference is WASH-1400, 'The Reactor Safety Study', which was produced for the Nuclear Regulatory Commission.
Upon investigating, I formed the impression that this case study little relevance to the question of whether we can know what to do about AI:
* The issue was one that a lot of people were already concerned about.
* The issue was highly domain specific.
* Rather than there being a few salient predictions, there were a huge number of small predictions.
One way in which the situation is relevant is that risk of nuclear power plant accidents was not adequately addressed, but that's only one of many examples people not adequately addressing risks.
A few details below.
* The report predicted one core damage accident per 20,000 years of nuclear power plant operation.
As a point of comparison, at the time of the Three Mile Island incident, only 500 years of nuclear power plant operation had occurred. This could be a fluke. The Chernobyl accident occurred only 6 years later. If the cause was the same, that strongly suggests that the Three Mile Island incident was not a fluke.
However, the report was based on reactor designs that didn't include the Three Mile Island type.
* The report did discuss tidal waves as a potential cause for nuclear disaster, anticipating the recent disaster in Japan. But the report is 21 volumes long, and so this (weak) prediction could have been cherry picked retrospectively.
* The report is considered to be obsolete, and its main lasting value se |
e2f71dde-1ad9-4664-a9de-987fe810e7aa | trentmkelly/LessWrong-43k | LessWrong | My Arrogant Plan for Alignment
Preface: Most of my predictions have great uncertainty when it comes to AI development and alignment, and I have great respect for the predictions of others.
However, I thought it might be fun and useful to write a plan based solely on predictions I would make if I assumed my predictions were more accurate than everyone else's (including betting markets).
Lastly, I point out why detailed plans for alignment are difficult and why broader and more flexible strategies are preferable.
Plan Summary:
1. Make 100M USD by founding an AI startup
2. Pivot into interpretability research and lobbying
3. Incentivize AI labs/politicians to aim for creating a "limited AGI" such as an oracle AGI that solves alignment for us, ideally in combination with using interpretability tools to detect deception
4. Use limited AGI to create aligned powerful ASI, most likely through emulating human brains.
5. Live happily ever after
Background - How the Paths I See Towards Aligning AGI
I can only think of three ways to create AGI with >95% probability of being aligned:
1. Have a bulletproof AI lie detection system, capable of telling if the AI in training is lying with extreme accuracy, and continuously check for alignment while developing ever smarter AI.
2. Whole brain emulations of humans, and improve the intelligence of those emulations
3. Training AIs on tasks so straightforward they will not do things outside of what they are optimized for, and then use that “constrained” AI to solve the alignment problem.
Plan Full
Step 1 - Making Money
Arrogant me: Money is a convergent goal, meaning that if I make money first, I can change my plan later on and still be in a good position. AI is likely the biggest ever opportunity to make loads of money.
AI requires technical expertise, which most big money decision makers lack and have to rely on others to help them make decisions. But since the decision makers lack my technical expertise, they can not pick technical advisors o |
0ff9c6df-ecc8-427c-917d-70fd4ce7cf91 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Branching Time Active Inference: empirical study and complexity class analysis
1 Introduction
---------------
Active inference extends the free energy principle to generative models with actions (Friston et al., [2016a](#bib.bib16); Da Costa et al., [2020](#bib.bib10); Champion et al., [2021b](#bib.bib7)) and can be regarded as a form of planning as inference (Botvinick and Toussaint, [2012](#bib.bib2)). This framework has successfully explained a wide range of brain phenomena, such as habit formation (Friston et al., [2016a](#bib.bib16)), Bayesian surprise (Itti and Baldi, [2009](#bib.bib22)), curiosity (Schwartenbeck et al., [2018](#bib.bib34)), and dopaminergic discharge (FitzGerald et al., [2015](#bib.bib11)). It has also been applied to a variety of tasks such as navigation in the Animal AI environment (Fountas et al., [2020](#bib.bib13)), robotic control (Pezzato et al., [2020](#bib.bib29); Sancaktar et al., [2020](#bib.bib32)), the mountain car problem (Catal et al., [2020](#bib.bib5)), the game DOOM (Cullen et al., [2018](#bib.bib9)) and the cart-pole problem (Millidge, [2019](#bib.bib24)).
Active inference builds on a subfield of Bayesian statistics called variational inference (Fox and Roberts, [2012](#bib.bib14)), in which the true posterior is approximated with a variational distribution. This method provides a way to balance the computational cost and accuracy of the posterior distribution. Indeed, the variational approach is only tractable because some statistical dependencies are ignored during the inference process, i.e., the variational distribution is generally assumed to fully factorise, leading to the well known mean-field approximation:
| | | |
| --- | --- | --- |
| | Q(X)=∏iQ(Xi) | |
where X is the set of all hidden variables of the model and Xi represents the i-th hidden variable. Winn and Bishop ([2005](#bib.bib40)) presented a message-based implementation of variational inference, naturally called variational message passing. And more recently, Champion et al. ([2021b](#bib.bib7)) rigorously framed active inference as a variational message passing procedure. By combining the Forney factor graph formalism (Forney, [2001](#bib.bib12)) with the method of Winn and Bishop ([2005](#bib.bib40)), it becomes possible to create modular implementations of active inference (van de Laar and de Vries, [2019](#bib.bib37); Cox et al., [2019](#bib.bib8)) that allows the users to define their own generative models without the burden of deriving the update equations. This paper uses a new software package called Homing Pigeon that implements such a modular approach and the relevant code has been made publicly available on github: <https://github.com/ChampiB/Homing-Pigeon.>
Arguably, the major bottleneck for scaling up the active inference framework was the exponential growth of the number of policies. In the reinforcement learning literature, this explosion is frequently handled using Monte Carlo tree search (MCTS) (Silver et al., [2016](#bib.bib35); Browne et al., [2012](#bib.bib4); Schrittwieser et al., [2019](#bib.bib33)). MCTS is based on the upper confidence bound for trees (UCT), which originally comes from the multi-armed bandit problem, and trades-off exploration and exploitation during the tree search. In the reinforcement learning litterature, the selection of the node to expand is carried out using the UCT criterion111This version of UCT comes from Silver et al. ([2016](#bib.bib35)), which is defined as:
| | | | |
| --- | --- | --- | --- |
| | UCT(s,a)=q(s,a)+CexploreP(s,a)1+N(s,a), | | (1) |
where q(s,a) is the value of taking action a in state s (i.e. q here is not the variational posterior), Cexplore is the exploration constant that modulates the amount of exploration, N(s,a) is the visit count, and P(s,a) is the prior probability of selecting action a in state s. MCTS has been applied to active inference in a recent paper by Fountas et al. ([2020](#bib.bib13)). The main change made to apply MCTS to active inference was to modify the node selection step that returns the node to be expanded. From equation (9) of (Fountas et al., [2020](#bib.bib13)), one can see that the UCT formula has been replaced by:
| | | | |
| --- | --- | --- | --- |
| | U(s,a)=−~G(s,a)+CexploreQ(a|s)1+N(s,a) | | (2) |
where ~G(s,a) is the best estimation of the expected free energy (EFE) obtained from Equation 8 of Fountas et al. ([2020](#bib.bib13)) using sampling of 3 (out of 4) neural networks used by the system, N(s,a) is the number of times that a was explored in state s, Cexplore is an exploration constant and Q(a|s) is a neural network modelling the posterior distribution over actions. Note, Q(a|s) specializes P(s,a) in equation ([1](#S1.E1 "(1) ‣ 1 Introduction ‣ Branching Time Active Inference empirical study and complexity class analysis")), by providing the probability of selecting action a in state s. One can see that U(s,a) has been obtained from UCT by replacing the average reward by the negative EFE.
More recently, Champion et al. ([2021a](#bib.bib6)) proposed an online method that frames planning as a form of structure learning guided by the expected free energy. This method, called branching-time active inference (BTAI), generalises active inference (Friston et al., [2016a](#bib.bib16); Champion et al., [2021b](#bib.bib7); Da Costa et al., [2020](#bib.bib10)) and relates to another recently introduced framework for inference and decision making, called sophisticated inference (Friston et al., [2021](#bib.bib15)). Importantly, the generative model of BTAI enables the agent to trade off risk and ambiguity, instead of only seeking for certainty as was the case in (Champion et al., [2021b](#bib.bib7)). In this paper, we provide an empirical study of BTAI, enabling us to explicitly demonstrate that BTAI provides a more scalable realization of planning as inference than active inference.
Section [2](#S2 "2 Branching Time Active Inference (BTAI) ‣ Branching Time Active Inference empirical study and complexity class analysis") reviews the BTAI theory, with full details presented in (Champion et al., [2021a](#bib.bib6)). Then, Section [3](#S3 "3 BTAI vs active inference ‣ Branching Time Active Inference empirical study and complexity class analysis") compares BTAI to standard active inference in the context of a graph navigation task both empirically and theoretically. We show that active inference is able to solve small graphs but suffers from an exponential (space and time) complexity class that makes the approach intractable for bigger graphs. In contrast, BTAI is able to search the space of policies efficiently and scale to bigger graphs. Next, Section [4.2](#S4.SS2 "4.2 Overcoming the challenge of local minima ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis") presents the challenge of local minima in the context of a maze solving task, and shows how better prior preferences and deeper tree search help to overcome this challenge. Lastly, Section [4.3](#S4.SS3 "4.3 Solving more mazes ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis") compares two cost functions, gclassic and gpcost, in two new mazes. Finally, Section [5](#S5 "5 Conclusion and future works ‣ Branching Time Active Inference empirical study and complexity class analysis") concludes this paper, and provides ideas for future research.
2 Branching Time Active Inference (BTAI)
-----------------------------------------
In this section, we provide a short review of BTAI, and the reader is referred to (Champion et al., [2021a](#bib.bib6)) for details. BTAI frames planning as a form of structure learning (Smith et al., [2020](#bib.bib36); Friston et al., [2016b](#bib.bib19), [2018](#bib.bib17)) guided by the expected free energy. The idea is to define a generative model that can be expanded dynamically as shown in Figure [1](#S2.F1 "Figure 1 ‣ 2 Branching Time Active Inference (BTAI) ‣ Branching Time Active Inference empirical study and complexity class analysis").
Figure 1:
This figure illustrates the expandable generative model allowing planning under active inference. The current time point (the present) is denoted by t. All times before t are the past, and after t are the future. States in the future are indexed by (multi-index) action sequences, with each digit indicating an action, e.g. S(11). The future is a tree-like generative model whose branches correspond to the policies considered by the agent. The branches can be dynamically expanded during planning and the nodes in light gray represent possible expansions of the current generative model.
The past and present is modelled using a partially observable Markov decision process (POMDP) in which each observation (Oτ) only depends on the state at time τ, and this state (Sτ) only depends on the previous state (Sτ−1) and previous action (Uτ−1). In addition to the POMDP which models the past and present, the future is modelled using a tree-like generative model whose branches are dynamically expanded. Each branch of the tree corresponds to a trajectory of states reached under a specific policy. The branches are expanded following a logic similar to the Monte Carlo tree search algorithm, and the state estimation is performed using variational message passing.
At the start of a trial, the model contains only the initial hidden state S0 and the initial observation O0. Then, the agent starts expanding the generative model using an approach inspired by Monte Carlo tree search (Browne et al., [2012](#bib.bib4)), where the selection of a node is based on expected free energy. The expansion of a node SI adds the future hidden state SI::U for each available action U, where I::U refers to the multi-index obtained by adding the action U at the end of the sequence of actions described by the multi-index I. The expansion step also adds the latent variable corresponding to observation OI::U to the generative model. Next, the evaluation step estimates the cost attached to the pairs (SI::U,OI::U) for each action U. In this paper, we experimented with two kinds of cost. First, the standard expected free energy:
| | | |
| --- | --- | --- |
| | gclassicJ\ensurestackMath\stackon[1pt]=ΔDKL[Q(OJ)||V(OJ)]+EQ(SJ)[H[P(OJ|SJ)]], | |
where J=I::U for an arbitrary action U, and gclassicJ trades off risk and ambiguity. Second, we also experiment with the following quantity:
| | | |
| --- | --- | --- |
| | gpcostJ\ensurestackMath\stackon[1pt]=ΔDKL[Q(SJ)||V(SJ)]+DKL[Q(OJ)||V(OJ)], | |
which depends on both the risk over observations and the risk over states. Lastly, the cost of the best action (i.e., the action that produces the smallest cost) is propagated towards the root and used to update the aggregated cost of the ancestors of SI.
Finally, during the planning procedure, the agent needs to perform inference of the future hidden states and observations. This is performed using variational message passing on the set of newly expanded nodes, i.e. {SI::U,OI::U∣U∈{1,...,|U|}}, until convergence to a minimum in the free energy landscape.
We refer the interested reader to (Champion et al., [2021b](#bib.bib7)) for additional information about the derivation of the update equations. Also, since this paper only considers inference and not learning (i.e. the model does not have Dirichlet priors over the tensors defining the world’s contingencies), the generative model is different from the one presented in the theoretical paper (Champion et al., [2021a](#bib.bib6)). Therefore, we provide a mathematical description of the generative model, the variational distribution and the belief updates in Appendix A.
3 BTAI vs active inference
---------------------------
In this section, we benchmark BTAI against standard active inference as implemented in Statistical Parametric Mapping (SPM), c.f. Friston ([2007](#bib.bib18)) for additional details about SPM. First, we do this in terms of complexity class and then empirically through experiments of increasing difficulty.
###
3.1 The deep reward environment
First, we introduce a canonical example of the kind of environment in which BTAI outperforms standard active inference. This environment is called the deep reward environment because the agent needs to navigate a tree like graph where the graph’s nodes correspond to the states of the system, and the agent needs to look deep into the future to diferentiate the good path from the traps.
At the beginning of each trial, the agent is placed at the root of the tree, i.e., the initial state of the system. From the initial states, the agent can perform n+m actions, where n and m are the number of good and bad paths, respectively. Additionally, at any point in time, the agent can make two observations: a pleasant one or an unpleasant one. The states of the good paths produce pleasant observations, while the states of the bad paths produces unpleasant ones.
If the first action selected was one of the m bad actions, then the agent will enter a bad path in which n+m actions are available at each time step but all of them produce unpleasant observations. If the first action selected was one of the n good actions, then the agent will enter the associated good path. We let Lk be the length of the k-th good path. Once the agent is engaged on the k-th path, there are still n+m actions available but only one of them keeps the agent on the good path. All the other actions will produce unpleasant observations, i.e., the agent will enter a bad path.
Figure 2:
This figure illustrates a type of environment in which BTAI will outperform standard active inference. Typically, this corresponds to environments in which there are only a small number of good actions. In such environments BTAI can safely discard a large part of the tree, and speed up the search without impacting performance. Note, S0 represents the initial state, Sb represents a bad state, Sg represents a good state, and Sij is the j-th state of the i-th good path. Also, the longest path in the above picture is the first good path whose length L1 is equal to two. Importantly, the longest path corresponds to the only good path that does not turn out to be a trap. Finally, at each time point, the agent must pick from the m+n possible actions. When reaching Sn1, the agent reaches the end of the n-th good path, and therefore the m+n possible actions all lead to an unpleasant state. When reaching S11, the agent is still able to continue on the first good path, but all the other actions (i.e. m+n−1 actions) lead to a bad state.
This process will continue until the agent reaches the end of the k-th path which is determined by the path’s length Lk. If the k-th path was the longest of the n good paths, then the agent will from now on only receive pleasant observations independently of the action performed. If the k-th path was not the longest path, then independently of the action performed, the agent will enter a bad path.
To summarize, at the beginning of each trial, the agent is prompted with n good paths and m bad paths. Only the longest good path will be beneficial in the long term, the others are traps, which will ultimately lead the agent to a bad state. Figure [2](#S3.F2 "Figure 2 ‣ 3.1 The deep reward environment ‣ 3 BTAI vs active inference ‣ Branching Time Active Inference empirical study and complexity class analysis") illustrates this environment.
###
3.2 BTAI vs active inference: Space and Time complexity
In this section, we compare our model to the standard model of active inference (Friston et al., [2016a](#bib.bib16); Da Costa et al., [2020](#bib.bib10)). In the standard formulation, the implementation needs to store the parameter of the posterior over states sπτ for each policy and each time step. Therefore, assuming |U| possible actions, T time steps, |π|=|U|T policies, and |S| possible hidden state values, the space complexity class for storing the parameters of the posterior over hidden states is O(|π|×T×|S|). This corresponds to the number of parameters that needs to be stored, and it is a problem because |π| grows exponentially with the number of time steps. Additionally, performing inference on an exponential number of parameters will lead to an exponential time complexity class.
BTAI solves this problem by allowing only K expansions of the tree. In BTAI, we need to store |S| parameters for each time step in the past and present, and for each expansion, we only need to compute and store the parameters of the posterior over the hidden states corresponding to this expansion. Therefore, the time and space complexity class is O([K+t]×|S|), where t is the current time point. This is linear in the number of expansions. Now, the question is how many expansions are required to solve the task? Even if the task requires the tree to be fully expanded, then the complexity class of BTAI would be O([|U|T−t+t]×|S|). Figure [3](#S3.F3 "Figure 3 ‣ 3.2 BTAI vs active inference: Space and Time complexity ‣ 3 BTAI vs active inference ‣ Branching Time Active Inference empirical study and complexity class analysis") illustrates the difference between AI and BTAI in terms of the space complexity class, when BTAI performs a full expansion of the tree.
Figure 3:
This figure illustrates the difference between AI and BTAI in terms of space complexity class. The time goes from top to bottom, we assume two actions at each time step, t denotes the current time point, and each circle represents the storage of the |S| parameters required to store a categorical distribution of a hidden state. Black nodes represent the nodes that must be stored in BTAI (under a full expansion of the tree), while the red nodes represent AI’s extra costs of storage. This extra cost comes from the fact that in AI, one needs to store posterior beliefs for each time step and for each policy, while in BTAI, the tree allows us to compress the representation.
Additionally to the gain afforded by the structure of the tree, most practical applications can be solved by expanding only a small number of nodes (Silver et al., [2016](#bib.bib35); Schrittwieser et al., [2019](#bib.bib33)), which means that MCTS and BTAI approaches will be even more optimised than in Figure [3](#S3.F3 "Figure 3 ‣ 3.2 BTAI vs active inference: Space and Time complexity ‣ 3 BTAI vs active inference ‣ Branching Time Active Inference empirical study and complexity class analysis") because most branches will not be expanded.
One could argue that there is a trade off in the nature and extent of the information inferred by classic active inference and branching-time active inference. Specifically, classic active inference exhaustively represents and updates all possible policies, while branching-time active inference will typically only represent a small subset of the possible trajectories. These will typically be the more advantageous paths for the agent to pursue, with the less beneficial paths not represented at all. Indeed, the tree search is based on the expected free energy that favors policies that maximize information gain while realizing the prior preferences of the agent.
Additionally, the inference process can update the system’s understanding of past contingencies on the basis of new observations. As a result, the system can obtain more refined information about previous decisions, perhaps re-evaluating the optimality of these past decisions. Because classic active inference represents a larger space of policies, this re-evaluation could apply to more policies.
We also know that humans engage in counterfactual reasoning (Rafetseder et al., [2013](#bib.bib30)), which, in our planning context, could involve the entertainment and evaluation of alternative (non-selected) sequences of decisions. It may be that, because of the more exhaustive representation of possible trajectories, classic active inference can more efficiently engage in counterfactual reasoning. In contrast, branching-time active inference would require these alternative pasts to be generated “a fresh” for each counterfactual deliberation. In this sense, one might argue that there is a trade off: branching-time active inference provides considerably more efficient planning to attain current goals, classic active inference provides a more exhaustive assessment of paths not taken.
###
3.3 BTAI vs active inference: Simulations
In this section, we compare BTAI to standard active inference in the context of three deep reward environments. All the environments have two good paths and five bad paths. However, the lengths of the two good paths (i.e., L1 and L2) changes from environment to environment as summarised in Table [1](#S3.T1 "Table 1 ‣ 3.3 BTAI vs active inference: Simulations ‣ 3 BTAI vs active inference ‣ Branching Time Active Inference empirical study and complexity class analysis"). For all the environments L2>L1, therefore the first path is a trap that will lead to a bad state, and the second path is the path the agent should take. Also, to identify that the first path is a trap, the agent must be able to plan at least L1+1 steps ahead, since before that the two good paths are identical.
| Environment | L1 | L2 |
| --- | --- | --- |
| first | 2 | 3 |
| second | 4 | 5 |
| third | 7 | 9 |
Table 1: This table presents the three deep reward environments on which experiments will be run.
Table [2](#S3.T2 "Table 2 ‣ 3.3 BTAI vs active inference: Simulations ‣ 3 BTAI vs active inference ‣ Branching Time Active Inference empirical study and complexity class analysis") shows the result of SPM simulations in which a standard active inference agent is run on the three deep reward environments presented in Table [1](#S3.T1 "Table 1 ‣ 3.3 BTAI vs active inference: Simulations ‣ 3 BTAI vs active inference ‣ Branching Time Active Inference empirical study and complexity class analysis"). As expected, the agent successfully solved the first two environments, for which it was required to plan three and five steps ahead. However, the third deep reward environment is intractable using standard active inference, i.e., the simulation runs out of memory because of the exponential (space) complexity class.
| | | | |
| --- | --- | --- | --- |
| Environment | Policy size | P(goal) | P(trap) |
| first | 3 | 1 | 0 |
| second | 5 | 1 | 0 |
| third | 8 | crash | crash |
Table 2: This table shows that the active inference agent was able to plan three and five time steps ahead to solve the first and the second deep reward environments. However, because of the exponential space complexity, SPM runs out of memory when trying to plan eight time steps ahead to solve the third deep reward environment.
Table [3](#S3.T3 "Table 3 ‣ 3.3 BTAI vs active inference: Simulations ‣ 3 BTAI vs active inference ‣ Branching Time Active Inference empirical study and complexity class analysis") shows the result obtained by BTAI on the three deep reward environments presented in Table [1](#S3.T1 "Table 1 ‣ 3.3 BTAI vs active inference: Simulations ‣ 3 BTAI vs active inference ‣ Branching Time Active Inference empirical study and complexity class analysis"). As expected, the agent successfully solved the three deep reward environments. Ten planning iterations were required for the first environment, fifteen were required for the second environment, and twenty for the third.
| | | | |
| --- | --- | --- | --- |
| Environment | Planning iterations | P(goal) | P(trap) |
| first | 10 | 1 | 0 |
| second | 10 | 0.49 | 0.51 |
| | 15 | 1 | 0 |
| third | 10 | 0.47 | 0.53 |
| | 15 | 0.55 | 0.45 |
| | 20 | 1 | 0 |
Table 3: This table shows that BTAI was able to solve the three deep reward environments with at most 20 planning iterations.
4 BTAI Empirical Intuition
---------------------------
In this section, we study the BTAI agent’s behaviour through experiments highlighting its vulnerability to local minimum and ways to mitigate this issue. The goal is to gain some intuition about how the model behaves when: enabling deeper searches, providing better preferences, and using different kind of cost functions to guide the Monte Carlo tree search. The code of those experiments is available on GitHub at the following URL: <https://github.com/ChampiB/Experiments_AI_TS.>
###
4.1 The maze environment
This section presents the environment in which various simulations will be run. In this environment, the agent can be understood as a rat navigating a maze. Figure [4](#S4.F4 "Figure 4 ‣ 4.1 The maze environment ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis") illustrates the three mazes studied in the following sections. The agent can perform five actions, i.e., UP, DOWN, LEFT, RIGHT and IDLE. The goal is to reach the maze exit from the starting position of the agent. To do so, the agent must move from empty cells to empty cells avoiding walls. If the agent tries to move through a wall, the action becomes equivalent to IDLE. Finally, the observations made by the agent correspond to the Manhattan distance (with the ability to traverse walls) between its current position and the maze exit. Taking the first maze of Figure [4](#S4.F4 "Figure 4 ‣ 4.1 The maze environment ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis") as an example, if the agent stands on the exit, the observation will be zero or equivalently using one hot encoding [1 0 0 0 0 0 0 0 0], and if the agent stands at the initial position, the observation will be nine or equivalently [0 0 0 0 0 0 0 0 1].
Figure 4:
This figure illustrates the three mazes used to perform the experiments in the next sections. We will refer to the leftmost maze as the first, the maze at the center as the second, and the rightmost maze as the third. Also black squares correspond to walls, green squares correspond to the maze exit and red squares correspond to the agent starting position.
###
4.2 Overcoming the challenge of local minima
In this section, we investigate the challenge of local minima and provide two ways of mitigating the issue: improving the prior preferences and using a deeper tree.
####
4.2.1 Prior preferences and local minimum
In this first experiment, the agent was asked to solve the second maze of Figure [4](#S4.F4 "Figure 4 ‣ 4.1 The maze environment ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis"), which has the property that the start location (red square) is a local minimum. Remember from Section [4.1](#S4.SS1 "4.1 The maze environment ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis") that the agent observes the Manhattan distance between his location and the maze exit. The Manhattan distance naturally creates local minima throughout the mazes, i.e., cells of the maze (apart from the exit) for which no adjacent cell has a lower distance to the exit. An example of such a local minimum is shown as a blue square in Figure [5](#S4.F5 "Figure 5 ‣ 4.2.1 Prior preferences and local minimum ‣ 4.2 Overcoming the challenge of local minima ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis"). The presence of such a local minimum implies that a well behaved agent (i.e., an agent trying to get as close as possible to the exit) might get trapped in those cells for which no adjacent cell has a lower distance to the exit and thus fail to solve the task.
Figure 5:
This figure illustrates the notion of local minimum (i.e., the blue cell) in the context of the first maze. Local minima correspond to cells (apart from the exit) for which no adjacent cell has a lower distance to the exit.
Next, we need to define the prior preferences of the agent. Our framework allows the modeller to define prior preferences over both future observations and future states. However, we start by assuming no preferences over the hidden states, i.e., V(SI) is uniform. We define the prior preferences over future observations as:
| | | |
| --- | --- | --- |
| | CO=σ(−γv) with v=[12...|O|]T | |
where |O| is the number of possible observations (9 in the first maze of Figure [4](#S4.F4 "Figure 4 ‣ 4.1 The maze environment ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis")), γ is the precision of the prior preferences, and σ(⋅) is the softmax function. The above prior preferences will give high probability to cells close to the exit and will exhibit the local minimum behaviours previously mentioned. Indeed, the minus sign before γv makes the elements in CO big, when the corresponding elements in v are small, and vice versa.
Using these prior preferences, we ran 100 simulations in the second maze of Figure [4](#S4.F4 "Figure 4 ‣ 4.1 The maze environment ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis"). Each simulation was composed of 20 action-perception cycles. Note, the results might vary from simulation to simulation, because the actions performed in the environment are sampled from σ(−ωgN), where σ(\vbox\scalebox{.5}{∙}) is a softmax function, ω is the precision of action selection, g is a vector whose elements correspond to the cost of the root’s children (i.e. the children of St) and N is a vector whose elements correspond to the number of visits of the root’s children.
Table [4](#S4.T4 "Table 4 ‣ 4.2.1 Prior preferences and local minimum ‣ 4.2 Overcoming the challenge of local minima ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis") reports the frequency at which the agent reaches the exit when using: an exploration constant Cp of 2 in the UCT criterion and the expected free energy as the cost function. First, note that with only 10 planning iterations, the agent is unable to leave the initial position (i.e., it is trapped in the local minimum). But as the number of planning iterations is increased, the agent becomes able to foresee the benefits of leaving the local minimum.
| Planning iterations | P(exit) | P(local) |
| --- | --- | --- |
| 10 | 0 | 1 |
| 15 | 0.98 | 0.02 |
| 20 | 1 | 0 |
Table 4: This table presents the probability that the agent solves the maze, and the probability of the agent being stuck into the local minimum.
####
4.2.2 Improving prior preference to avoid local minimum
In this second experiment, we modified the prior preferences of the agent to enable it to avoid local minima. We first change the cost function from the expected free energy gclassicI to the pure cost gpcostI, which allows us to set nontrivial preferences over states (in the previous simulation, these were set to uniform). Specifically, the prior preferences over hidden states will be of the form:
| | | |
| --- | --- | --- |
| | CS=σ(−γw), | |
where γ is the precision over prior preferences, and w is set according to Figure [6](#S4.F6 "Figure 6 ‣ 4.2.2 Improving prior preference to avoid local minimum ‣ 4.2 Overcoming the challenge of local minima ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis"). Also, we set the precision γ to two and the exploration constant in the UCT criterion to 2.4. Finally, the prior preferences over future observations remain the same as in the previous section.
Figure 6:
This figure illustrates the new prior preferences of the agent over the future states. Black squares correspond to walls, the darkest red corresponds to high prior preferences (really enjoyable states), the brightest red corresponds to low prior preferences (annoying states) and the last kind of red corresponds to medium prior preferences (boring states).
Tables [5](#S4.T5 "Table 5 ‣ 4.2.2 Improving prior preference to avoid local minimum ‣ 4.2 Overcoming the challenge of local minima ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis") and [6](#S4.T6 "Table 6 ‣ 4.2.2 Improving prior preference to avoid local minimum ‣ 4.2 Overcoming the challenge of local minima ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis") summarize the results of the experiments with and without the use of prior preferences over hidden states, respectively. As expected better prior preferences lead to better performance when less planning iterations are performed. Specifying prior preferences over hidden states requires the modeller to bring additional knowledge to the agent, and might not always be possible. However, when such knowledge is available it can improve the agent’s performance. This illustrates the value of the BTAI approach, which enables preferences to be specified for observations, as does active inference, as well as for states.
| | | |
| --- | --- | --- |
| Planning iterations | P(global) | P(local) |
| 10 | 0 | 1 |
| 15 | 0 | 1 |
| 20 | 1 | 0 |
Table 5: This table presents the probability that the agent solves the maze, and the probability of the agent being stuck in the local minimum. In this table, the agent was not equipped with prior preferences over hidden states.
| | | |
| --- | --- | --- |
| Planning iterations | P(global) | P(local) |
| 10 | 0 | 1 |
| 15 | 1 | 0 |
| 20 | 1 | 0 |
Table 6: This table presents the probability that the agent solves the maze, and the probability of the agent being stuck in the local minimum. In this table, the agent was equipped with prior preferences over hidden states.
###
4.3 Solving more mazes
Up to now, we focused on the second maze of Figure [4](#S4.F4 "Figure 4 ‣ 4.1 The maze environment ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis") to demonstrate that both improving prior preferences and deepening the tree can help to mitigate the problem of local minima. In this section, we extend our analysis to the first and third mazes. Tables [7](#S4.T7 "Table 7 ‣ 4.3 Solving more mazes ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis") and [8](#S4.T8 "Table 8 ‣ 4.3 Solving more mazes ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis") shows the performance of the BTAI agent in the first maze of Figure [4](#S4.F4 "Figure 4 ‣ 4.1 The maze environment ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis") when using gclassicI and gpcostI as cost function, respectively. When gpcostI was used as a cost function, the agent was only equipped with prior preferences over observations (i.e., uniform preferences over hidden states). Additionally, in both tables, the exploration constant of the UCT criterion was set to two.
| | | |
| --- | --- | --- |
| Planning iterations | P(global) | P(local) |
| 10 | 1 | 0 |
| 15 | 1 | 0 |
| 20 | 1 | 0 |
Table 7: This table presents the probability that the agent solves the first maze of Figure [4](#S4.F4 "Figure 4 ‣ 4.1 The maze environment ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis"), and the probability of the agent falling into the local minimum. In this table, the agent was using gclassicI as a cost function.
| | | |
| --- | --- | --- |
| Planning iterations | P(global) | P(local) |
| 10 | 0 | 1 |
| 15 | 1 | 0 |
| 20 | 1 | 0 |
Table 8: This table presents the probability that the agent solves the first maze of Figure [4](#S4.F4 "Figure 4 ‣ 4.1 The maze environment ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis"), and the probability of the agent falling into the local minimum. In this table, the agent was using gpcostI as a cost function.
Importantly, BTAI is able to solve the task using both cost functions. Also, gclassicI performs better than gpcostI on this maze. However, this finding does not generalize to the third maze of Figure [4](#S4.F4 "Figure 4 ‣ 4.1 The maze environment ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis"), where both gclassicI and gpcostI lead to the same results when the precision over prior preference was set to one, and the exploration constant of the UCT criterion was set to two. Those results are presented in Table [9](#S4.T9 "Table 9 ‣ 4.3 Solving more mazes ‣ 4 BTAI Empirical Intuition ‣ Branching Time Active Inference empirical study and complexity class analysis").
| | | |
| --- | --- | --- |
| Planning iterations | P(global) | P(local) |
| 10 | 0 | 1 |
| 15 | 0 | 1 |
| 20 | 1 | 0 |
Table 9: This table presents the probability that the agent solves the maze, and the probability of the agent falling into the local minimum. Both cost functions gclassicI and gpcostI lead to the above results in the third maze.
5 Conclusion and future works
------------------------------
In this paper, we provided an empirical study of branching time active inference (BTAI), where the name takes inspiration from branching-time theories of concurrent and distributed systems in computer science (Glabbeek, [1990](#bib.bib20); van Glabbeek, [1993](#bib.bib38); Bowman, [2005](#bib.bib3)), and planning was cast as structure learning (Smith et al., [2020](#bib.bib36); Friston et al., [2016b](#bib.bib19), [2018](#bib.bib17)). Simply put, the generative model is dynamically expanded and each expansion leads to the exploration of new policy fragments. The expansions are guided by the expected free energy, which provides a trade off between exploration and exploitation. Importantly, this approach is composed of not two, but three major distributions. The first is the prior distribution (or generative model) that encodes the agent’s beliefs before performing any observations. The second is the posterior (or variational) distribution encoding the updated beliefs of the agent after performing some observations. And the third is a target distribution over future states and observations that encodes the prior preferences of the agent, i.e., a generalization of the C matrix in the standard formulation of active inference proposed by Friston et al. ([2016a](#bib.bib16)). An important advantage of this generalization is that it allows the specification of prior preferences over both future observations and future states at the same time.
We compared BTAI and standard active inference theoretically by studying its space and time complexity class. This study highlights that our method should perform better than the standard model used in active inference when the task can be solved by expanding the tree only a small number of times with respect to an exhaustive search. Second, we compared BTAI to active inference empirically within the deep reward environment. Those simulations suggest that BTAI is able to solve problems for which a standard active inference agent would run out of memory. As elaborated upon in Section [3.2](#S3.SS2 "3.2 BTAI vs active inference: Space and Time complexity ‣ 3 BTAI vs active inference ‣ Branching Time Active Inference empirical study and complexity class analysis"), one might argue that there is a trade off between banching-time active inference, which provides considerably more efficient planning to attain current goals, and classic active inference which provides a more exhaustive assessment of paths not taken. This might enable active inference to more exhaustively reflect counter-factuals and reasoning based upon them.
Also, BTAI was studied (experimentally) in the context of a maze solving task and we showed that when the heuristic used to create the prior preferences is not perfect, the agent becomes vulnerable to local minima. In other words, the agent might be attracted by a part of the maze that has low cost but does not allow it to solve the task. Then, we demonstrated empirically that improving the prior preferences of the agent by specifying a good prior over future hidden states and deepening the tree search, helped to mitigate this issue.
This paper could lead to a large number of future research directions. One could for example add the ability of the agent to learn the transition matrices B as well as the likelihood matrix A and the vector of initial states D. This can be done in at least two ways. The first is to add Dirichlet priors over those matrices/vectors and the second would be to use neural networks as function approximators. The second option will lead to a deep active inference agent (Sancaktar and Lanillos, [2020](#bib.bib31); Millidge, [2020](#bib.bib25)) equiped with tree search that could be directly compared to the method of Fountas et al. ([2020](#bib.bib13)). Including deep neural networks in the framework will also open the door to direct comparison with the deep reinforcement learning literature (Haarnoja et al., [2018](#bib.bib21); Mnih et al., [2013](#bib.bib27); van Hasselt et al., [2015](#bib.bib39); Lample and Chaplot, [2017](#bib.bib23); Silver et al., [2016](#bib.bib35)). Those comparisons will enable the study of the impact of the epistemic terms when the agent is composed of deep neural networks.
Another direction of research will be to set up behavioural experiments to try to determine which kind of planning is used by the brain. This could simply be done by looking at the time required by a human to solve various mazes and compare it with both the classic model and the tree search alternative. Finally, one could also set up a hierarchical model of action and compare it to the tree search algorithm presented here. One could also evaluate the plausibility of a hierarchical model of action by running behavioural experiments on humans.
Finally, a completely different direction will be to focus on the integration of memory. At the moment, when a new action is performed in the environment and a new observation is recieved from it, all the branches in the tree are prunned and a new temporal slice (i.e. a new state, action and observation triple) is added to the POMDP. In other words, the integration function simply records the past. This exact recording of the past is very unlikely to really happen in the brain. Therefore, one might simply ask what to do with this currently ever growing record of the past. This would certainly lead to the notion of an active inference agent equipped with episodic memory (Botvinick et al., [2019](#bib.bib1)). |
fe680377-16d7-4c9e-86f2-f176a7b946ef | trentmkelly/LessWrong-43k | LessWrong | How can I reconcile the two most likely requirements for humanities near-term survival.
1. We technologically plateau, due to humanities questionable ability to adapt to accelerating technological progression.
2. AI development is indefinitely disrupted; as it is likely to result in disaster.
* This is unlikely to be done deliberately. Ongoing attempts to slow down AI development are relatively ineffective; it is more likely, in my opinion, that a basic form of AI developed in the near future will either directly increase all individuals power, or lead to technologies that do the same. An example would be the possible implications of Palm AI. https://www.theatlantic.com/technology/archive/2022/06/google-palm-ai-artificial-consciousness/661329/
* This universal increase in power could be sufficient to disrupt all AI research indefinitely, with the actions of a minority working in unison.
All this considered, what's the most likely situation that could play out? |
648b9b03-bd8e-454a-8c1e-e4601fc8ddea | StampyAI/alignment-research-dataset/lesswrong | LessWrong | (notes on) Policy Desiderata for Superintelligent AI: A Vector Field Approach
***Meta:** I thought I'd spend a little time reading the policy papers that Nick Bostrom has written. I made notes as I went along, so I spent a little while cleaning them up into a summary post. These are my notes on Bostrom, Dafoe and Flynn's 2016 [policy desiderata paper](https://nickbostrom.com/papers/aipolicy.pdf), which received significant edits in 2018. I spent 6-8 hours on this post, not a great deal of time, so I've not been maximally careful.*
Context and Goals
=================
Overall, this is not a policy *proposal*. Nor does it commit strongly to a particular moral or political worldview. The goal of this paper is to merely observe which policy challenges are especially important or different in the case of superintelligent AI, that most moral and political worldviews will need to deal with. The paper also makes no positive argument for the importance or likelihood or timeline of superintelligent AI - it instead assumes that this shall occur in the present century, and then explores the policy challenges that would follow.
The Vector Field Approach
-------------------------
Botrom, Dafoe and Flynn spend a fair amount of time explaining that they’re not going to be engaging in what (I think) Robin Hanson would call [standard value talk](http://www.overcomingbias.com/2013/10/beware-value-talk.html). They’re not going to endorse a particular moral or political theory, nor are they going to adopt various moral or political theories and show how they propose different policies. They’re going to look at the details of this particular policy landscape and try to talk about the regularities that will need to be addressed by most standard moral and political frameworks, and in what direction these regularities suggest changing policy.
They call this the ‘vector field’ approach. If you don't feel like you fully grok the concept, here's the quote where they lay out the formalism (with light editing for readability).
> The vector field approach might then attempt to derive directional policy change conclusions of a form that we might schematically represent as follows:
> “However much emphasis .mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
> .MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
> .mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
> .mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
> .mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
> .mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
> .mjx-numerator {display: block; text-align: center}
> .mjx-denominator {display: block; text-align: center}
> .MJXc-stacked {height: 0; position: relative}
> .MJXc-stacked > \* {position: absolute}
> .MJXc-bevelled > \* {display: inline-block}
> .mjx-stack {display: inline-block}
> .mjx-op {display: block}
> .mjx-under {display: table-cell}
> .mjx-over {display: block}
> .mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
> .mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
> .mjx-stack > .mjx-sup {display: block}
> .mjx-stack > .mjx-sub {display: block}
> .mjx-prestack > .mjx-presup {display: block}
> .mjx-prestack > .mjx-presub {display: block}
> .mjx-delim-h > .mjx-char {display: inline-block}
> .mjx-surd {vertical-align: top}
> .mjx-mphantom \* {visibility: hidden}
> .mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
> .mjx-annotation-xml {line-height: normal}
> .mjx-menclose > svg {fill: none; stroke: currentColor}
> .mjx-mtr {display: table-row}
> .mjx-mlabeledtr {display: table-row}
> .mjx-mtd {display: table-cell; text-align: center}
> .mjx-label {display: table-row}
> .mjx-box {display: inline-block}
> .mjx-block {display: block}
> .mjx-span {display: inline}
> .mjx-char {display: block; white-space: pre}
> .mjx-itable {display: inline-table; width: auto}
> .mjx-row {display: table-row}
> .mjx-cell {display: table-cell}
> .mjx-table {display: table; width: 100%}
> .mjx-line {display: block; height: 0}
> .mjx-strut {width: 0; padding-top: 1em}
> .mjx-vsize {width: 0}
> .MJXc-space1 {margin-left: .167em}
> .MJXc-space2 {margin-left: .222em}
> .MJXc-space3 {margin-left: .278em}
> .mjx-test.mjx-test-display {display: table!important}
> .mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
> .mjx-test.mjx-test-default {display: block!important; clear: both}
> .mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
> .mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
> .mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
> .mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
> .MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
> .MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
> .MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
> .MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
> .MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
> .MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
> .MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
> .MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
> .MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
> .MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
> .MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
> .MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
> .MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
> .MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
> .MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
> .MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
> .MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
> .MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
> .MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
> .MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
> .MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
> .MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
> .MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
> .MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
> .MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
> @font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
> @font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
> @font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
> @font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
> @font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
> @font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
> @font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
> @font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
> @font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
> @font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
> @font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
> @font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
> @font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
> @font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
> @font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
> @font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
> @font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
> @font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
> @font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
> @font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
> @font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
> @font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
> @font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
> @font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
> @font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
> @font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
> @font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
> @font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
> @font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
> @font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
> @font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
> @font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
> X you think that states ought, under present circumstances, to give to the objective of economic equality, there are certain special circumstances Y, which can be expected to hold in the radical AI context we described above, that should make you think that in those circumstances states should instead give emphasis fY(X) to the objective of economic equality."
> The idea is that f here is some relatively simple function, defined over a space of possible evaluative standards or ideological positions. For instance, f might simply add a term to X, which would correspond to the claim the emphasis given economic equality should be increased by a certain amount in the circumstances Y (according to all the ideological positions under consideration).
> Or f might require telling a more complicated story, perhaps along the lines of:
> “However much emphasis you give to economic equality as a policy objective under present circumstances, under conditions Y you should want to conceive of economic equality differently—certain dimensions of economic inequality are likely to become irrelevant and other dimensions are likely to become more important or policy-relevant than they are today.”
I particularly like this quote:
> This vector field approach is only fruitful to the extent that there are some patterns in how the special circumstances Y impact policy assessments from different evaluative positions. If the prospect of radical AI had entirely different and idiosyncratic implications for every particular ideology or interest platform, then the function f would amount to nothing more than a lookup table.
I read this as saying something like “This paper only makes sense if *facts* matter, separate to *values*.” It’s funny to me that this sentence felt necessary to be written.
Quotes
------
A few more quotes on what the paper is trying to do.
> A strong proposal for the governance of advanced AI would ideally accommodate each of these desiderata to a high degree. There may exist additional desiderata that we have not identified here; we make no claim that our list is complete. Furthermore, a strong policy proposal should presumably also integrate many other normative, prudential, and practical considerations that are either idiosyncratic to particular evaluative positions or are not distinctive to the context of radical AI.
> [...]
> Using a “vector field” approach to normative analysis, we sought to extract directional policy implications from these special circumstances. We characterized these implications as a set of desiderata—traits of future policies, governance structures, or decision-making contexts that would, by the standards of a wide range of key actors, stakeholders, and ethical views, enhance the prospects of beneficial outcomes in the transition to a machine intelligence era
> [...]
> By “policy proposals” we refer not only official government documents but also plans and options developed by private actors who take an interest in long-term AI developments. The desiderata, therefore, are also relevant to some corporations, research funders, academic or non-profit research centers, and various other organizations and individuals.
Next are the actual desiderata. They're given under four headings (efficiency, allocation, population, and process), each with 2-4 desiderata. Each subheading below corresponds to a policy desiderata in the paper. For each desiderata I have summarised of all the arguments and considerations in the text that felt new or non-trivial to me personally (e.g. I spent only one sentence on the arguments for AI safety).
If you want to just read the paper's summary, [jump down to page 23](https://nickbostrom.com/papers/aipolicy.pdf) which has a table and summarises in their own words.
Efficiency Desiderata
=====================
Expeditious progress
--------------------
We should make sure to take ahold of our cosmic endowment - and the sooner the better.
AI safety
---------
Choose policies that leads us to develop sufficient technical understanding that the AI will do what we expect it to do, and that give these tools to AI builders.
Conditional stabilization
-------------------------
The ability to establish a singleton, or regime of intensive global surveillance, or ability to thoroughly suppress the spread of dangerous or info, should we need to use this ability in the face of otherwise catastrophic global coordination failures.
Non-turbulence
--------------
Technology will change rapidly. We don’t want to have to rush regulations through, or alternatively take too long to adapt such that the environment radically changes again. So try to reduce turbulence.
Allocation Desiderata
=====================
Universal benefit
-----------------
If you force someone to take a risk, it is only fair that they are compensated with a share of any reward gained. Existential risks involve everyone, so everyone should get proportional benefit.
Epsilon-magnanimity
-------------------
Many people’s values have diminishing returns to further resources e.g. income guarantees for all, ensuring all animals have minimally positive lives, aesthetic projects like preserving some artworks, etc. While today they must fight for a cut of the small pie, as long as they are granted a non-zero weighting in the long-run, they can be satisfied. 0.00001% of GDP may be more than enough to give all humans a $40k income, for example.
This is especially good in light of normative uncertainty - as long as we give some weighting to various values, they will get satiated in a basic way in the long-run.
Continuity
----------
Reasons to expect unusually high concentration and permutation of wealth and power:
* In the modern world, salary is more evenly distributed than capital. Superintelligent AI is likely to greatly increase the factor share of income accrued from capital, leading to massive increases in inequality and increase concentration of wealth.
* If a small group decides how the AI works and its high-level decisions, they could gain a decisive strategic advantage and take over the world.
* If there is radical and unpredictable technological change, then it is likely that wealth distribution will change radically and unpredictably.
* Automated security and surveillance systems will help a regime stay alive without support from the public or elites - when behaviour is more legible it’s easier to punish or control it. This is also likely to at least sustain concentration of wealth and power, but also to increase it.
As such we wish to implement policies that more sustain existing concentration and distribution of wealth and power.
Also of interest, is (given the high likelihood of redistribution, change in concentration, and general unpredictable turbulence) how much we seem to face a global, real-life, Rawlsian veil-of-ignorance. It might be good to set up things like insurance to make sure everyone gets some minimum of power and self-determination in the future (it seems that people have diminishing returns to power - “most people would much rather be certain to have power over one life (their own) than have a 10% chance of having power over the lives of ten people and a 90% chance of having no power.”
Population Desiderata
=====================
Mind crime prevention
---------------------
Four key factors: novelty, invisibility, difference, and magnitude.
* Novelty and invisibility: Sentient digital entities may be moral patients. They would be a novel type of mind, and would not exhibit many characteristics that inform our moral intuitions - they lack facial expressions, physicality, human speech, and so on, if they are being run invisibly in some microprocessor. This means we should worry about policy makers taking an unconscionable moral decision.
* Difference: It is also the case that these minds may be very different to human or animal minds, again subverting our intuitions about what behaviour is normative toward them, and increasing the complexity of choosing sensible policies here.
* Magnitude: It may be incredibly cheap to create as many people as currently exist in a country, magnifying the concerns of the previous three factors. “With high computational speed or parallelization, a large amount of suffering could be generated in a small amount of wall clock time.” This may mean that mind crime is a principal desideratum in AI policy.
Population policy
-----------------
This is a worry about malthusian scenarios (where average income falls to subsistence levels). Hanson has written about these scenarios.
This can also undermine democracy (“One person, one vote”). If a political faction can invest in creating more people, they can create the biggest voting block. This leaves the following trilemma of options:
* (i) deny equal votes to all persons
* (ii) impose constraints on creating new persons
* (iii) accept that voting power becomes proportional to ability and willingness to pay to create voting surrogates, resulting in both economically inefficient spending on such surrogates and the political marginalization of those who lack resources or are unwilling to spend them on buying voting power
Some interesting forms of (i):
* Make voting rights something you inherit, a 1-1 mapping.
* Robin Hanson has suggested ‘speed-weighted voting’, because faster ems are more costly, so you'd actually have to pay a lot for marginal voters. This still looks like richer people getting a stronger vote, but in-principle puts a much higher cost on it.
Process Desiderata
==================
First principles thinking, wisdom, and technical understanding
--------------------------------------------------------------
Overall this is an especially different environment than usual policy-making, which means that we will need to be able to reconsider fundamental assumptions using first-principles thinking to a greater extent than before and be exceptionally wise (able to get the right answer to the most important questions while they are surrounded by confusion and misunderstanding).
Technological innovation is the primary driver of this radical new policy landscape, and so an understanding of the technologies is unusually helpful.
Speed and decisiveness
----------------------
In many possible futures, historic events will be happening faster than global treaties are typically negotiated, ratified, and implemented. We need a capacity for rapid decision-making and decisive global implementation.
Adaptability
------------
Many fundamental principles will need to be re-examined. Some examples: legitimacy, consent, political participation, accountability.
**Voluntary consent.** Given AIs that are super-persuaders and can convince anyone of anything, consent becomes a much vaguer and fuzzier concept. Perhaps consent only counts if the consentee has an “AI guardian” or “AI advisor” of some sort.
**Political participation.** This norm is typically justified on three grounds:
* Epistemic benefit of including information from a maximal diversity of sources.
* Ensures all interests and preferences are given some weighting in the decision.
* Intrinsic good.
However,
* The epistemic effect may become negative if the AI making decisions sits at a sufficiently high epistemic vantage point.
* AI may be able to construct a process / mechanism that accounts for all values without consistent input from humans.
* The intrinsic good is not changed, though it may not be worth the cost if the above to factors become strongly net negative and wasteful.
The above examples, of consent and political participation, are not at all clear, but just go to show that there are many unquestioned assumptions in modern political debate that may need either reformulation, abandonment, or extra vigilance spent on safeguarding their existence into the future.
Changes since 2016
==================
The paper was originally added to Nick Bostrom's website in 2016, and received an update in late 2018 ([original](https://www.fhi.ox.ac.uk/wp-content/uploads/Policy-Desiderata-in-the-Development-of-Machine-Superintelligence.pdf), [current](https://nickbostrom.com/papers/aipolicy.pdf)).
The main updates as I can see them are:
* The addition of 'vector field approach' to the title and body. It was lightly alluded to in the initial version. (I wonder if this was due to lots of feedback trying to fit the paper into standard value talk, where it did not want to be.)
* Changing the heading from "Mode" to "Process", and fleshing out the three desiderata rather than a single one called "Responsibility and wisdom". If you read the initial paper, this is the main section to re-read to get anything new.
There have definitely being significant re-writings of the opening section, and there may be more, but I did not take the time to compare them section-for-section.
*I've added some personal reflection/updates in a [comment](https://www.lesswrong.com/posts/hyfedqhgCQriBB9wT/notes-on-policy-desiderata-for-superintelligent-ai-a-vector#2YCHH7FoLuDS68RRR).* |
7606e702-64ae-4080-836c-b2510d4954dc | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | AI Safety Camp, Virtual Edition 2023
AI Safety Camp connects you with an experienced research lead to collaborate on a research project – helping you try your fit for a potential career in AI Safety research.
*The applications for AI Safety Camp’s Virtual Edition in 2023 are* [*now open*](https://airtable.com/shrNF1n4m1yqzCjbW)*!*
AI Safety Camp Virtual 8 will be a 3.5-month long online research program from 4 March to 18 June 2023, where participants form teams to work on pre-selected projects.
We value people with diverse backgrounds and skillsets, such as cybersecurity or cognitive science. Not all projects require participants to have prior experience in AI Safety, mathematics or machine learning. You will be able to read in detail about the research topics & each project’s skill requirements for our upcoming edition by following the Project Proposal links below.
Projects
========
Projects you can apply to…
Conceptualise AGI dynamics
--------------------------
[**Uncertainty -> Soft Optimization**](https://docs.google.com/document/d/1ayZ6goQhbvNZCSRXE7GFTSJm3Ptlenv7KjLr62Q3EzY/edit?usp=sharing) with Jeremy Gillen
[**Inquire into Uncontrollable Dynamics of AGI**](https://docs.google.com/document/d/1Zvt-tUksgtsuFs0LG8k6AHtDBmzHTOk-CfSz6nUYJrM/edit?usp=sharing) with Remmelt Ellen
[**Discussing and Crystallising a Research Agenda Based on Positive Attractors and Inherently Interpretable Architectures**](https://docs.google.com/document/d/1Fh2nCtNICVX60iFEwHHU6pZHZUoiKLX7InL_5e5dO9I/edit?usp=sharing) with Robert Kralisch
Investigate transformer models
------------------------------
[**Cyborgism**](https://docs.google.com/document/d/16YZljz2w44Fzqs0ZXgVKhkef8tTPkypjp7PK4S32SSs/edit?usp=sharing) with Nicholas Kees Dupuis
[**Understanding Search in Transformers**](https://hackmd.io/@YoMjTL9WTqOXv85V9skJxw/HJzTofMFj/edit) with Michael Ivanitskiy
[**Interdisciplinary Investigation of DebateGPT**](https://docs.google.com/document/d/1iDraPS4qst_-jcpxEpBN6xT0y80EWnTu02R7uAvQOno/edit) with Paul Bricman
Finetune language transformers
------------------------------
[**Does Introspective Truthfulness Generalize in LMs?**](https://docs.google.com/document/d/1qugeQWJwfwbu59ykxWwK95oMPi9aTZeQspMJQQvo-fo/edit?usp=sharing) with Jacob Pfau
[**Inducing Human-Like Biases in Moral Reasoning LMs**](https://docs.google.com/document/d/17GPZR289OVYVtBN6LfFdguz-tGO3fQh0x_48nMcevig/edit?usp=sharing) with Bogdan-Ionut Cirstea
Create annotation tools
-----------------------
[**Behavioral Annotation Framework for the Contextualized and Personalized Fine-Tuning of Foundation Models**](https://docs.google.com/document/d/1HP0tli2g6cQUme0mVQ4gfVQnFmt_IcN8YdpHWKRSC-g/edit?usp=sharing) with Eleanor “Nell” Watson
Review and analyse literature
-----------------------------
[**How Should Machines Learn from Default Options?**](https://docs.google.com/document/d/1xJm6yul6V-FSZOTEGxguJa99JEKKYMBPNnPrbCjB6RU/edit?usp=sharing) with En Qi Teo
[**Literature Review of the Neurological Basis of Human Values and Preferences**](https://docs.google.com/document/d/1YrqlM8i2SRj3MoojD4dKDv1f_-1Ezo0AyV3yDqEL-Pk/edit?usp=sharing) with Linda Linsefors
[**Machine learning for Scientific Discovery: the Present and Future of Science-Producing AI Models**](https://docs.google.com/document/d/1J6aadVK6b2nQGaVNPu0bV13ki0ijEy576CFGF0qV3oo/edit?usp=sharing) with Eleni Angelou
Propose public policy/communication
-----------------------------------
[**Policy Proposals for High-Risk AI Regulation**](https://docs.google.com/document/d/1FQGMjPBLWVIiuzFVBpVScqpi6nzeY0CWNK_Edd2F59M/edit?usp=sharing) with Koen Holtman
[**Developing Specific Failure Stories about Uncontrollable AI**](https://docs.google.com/document/d/1U--XV5Cuu5lug1kIJLJGYYEaqvYrsUZV7d1Q7cDR_ss/edit?usp=sharing) with Karl von Wendt
[**Apply to AISC Virtual 2023**](https://airtable.com/shrNF1n4m1yqzCjbW)
========================================================================
### Apply if you…
1. want to try out & consider ways you could help ensure that future AI performs safely and in line with what people value upon reflection;
2. are ready to dig into our research leads’ research topics and are able to write down a few clear arguments for why you’d research one in particular & how you might start;
3. previously studied a topic or practiced skills unique to your perspective/background that can bolster your new research team’s progress;
4. can block off hours to focus on research from **March to June 2023** on normal workdays and the weekends (at least 10 hours per week).
### Application timeline
| | | |
| --- | --- | --- |
| 5 Jan 2023 | 00:01 UTC | Project proposals are posted on our website. Application form [opens](https://airtable.com/shrNF1n4m1yqzCjbW). Reviews start right away . |
| 19 Jan | 23:59 AoE | Deadline to apply to teams closes. Late submissions might not get a response. |
| 1 March | 23:59 AoE | Last applicant admitted or declined (most will be informed of our decision earlier). |
*First virtual edition – a spontaneous collage*
Timeline
========
* January 5: Accepted proposals are posted on the AISC website. Application to join teams [open](https://airtable.com/shrNF1n4m1yqzCjbW).
* January 19: Application to join teams closes.
* Until February end: Organisers pre-filter applications. RLs interview potential members and pick their team.
* March 4-5: Opening weekend.
From here, teams meet weekly, and plan in their own work hours.
* June 17-18: Closing weekend.
All teams present their results.
Team structure
==============
Every team will have:
* one Research Lead
* one Team Coordinator
* other team members
All team members are expected to work at least 10 hours per week on the project, which includes joining weekly team meetings, and communicating regularly (between meetings) with other team members about their work.
As of yet, we cannot commit to offering stipends compensation for team members, because a confirmed grant fell through. Another grantmaker is in the midst of evaluating a replacement grant for AI Safety Camp. If confirmed, team members can opt in to receive a minimum of $500 gross per month (up to $2000 for full-time work).
Research Lead (RL)
------------------
The RL is the person behind the research proposal. If a group forms around their topics, the RL will guide the research project, and keep track of relevant milestones. When things inevitably don’t go as planned (this is research after all) the RL is in charge of setting the new course.
The RL is part of the research team and will be contributing to research the same as everyone else on the team.
Team Coordinator (TC)
---------------------
The TC is the ops person of the team. They are in charge of making sure meetings are scheduled, checks in with individuals on their task progress, etc. TC and RL can be the same person.
The role of the TC is important but not expected to take too much time (except for project management-heavy teams). Most of the time, the TC will act like a regular team member contributing to the research, same as everyone else on the team.
Other team members
------------------
Other team members will work on the project under the leadership of the RL and the TC. Team members will be selected based on relevant skills, understandings and commitments to contribute to the research project.
Questions?
==========
You can contact us at [contact@aisafety.camp](mailto:contact@aisafety.camp) |
2ffa4ee8-2446-4149-a74c-8e4822125658 | trentmkelly/LessWrong-43k | LessWrong | Two Agent Mild Optimization
|
dd9d7069-b98a-4665-9193-9a3ac6bd5e6e | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Asya Bergal: Reasons you might think human-level AI is unlikely to happen soon
---
*If we knew that human-level AI couldn’t be produced within the next 20 years, we would take different actions to improve the long-term future. Asya Bergal of* [*AI Impacts*](https://aiimpacts.org/) *talks about her investigations into why people say we won’t soon have human-level AI, including survey data, trends in “compute,” and arguments that current machine-learning techniques are insufficient.*
*We’ve lightly edited Asya’s talk for clarity. You can also watch it on* [*YouTube*](https://youtu.be/oFJcvyxpGSo) *and read it on* [*effectivealtruism.org*](https://effectivealtruism.org/articles/asya-bergal-reasons-you-might-think-human-level-ai-is-unlikely-to-happen)*.*
The Talk
--------
**Aaron Pang (Moderator):** Hello, and welcome to “Reasons you might think human-level AI is unlikely to happen soon” with Asya Bergal. Following a 20-minute talk by Asya, we will move to a live Q and A session where she will respond to your questions.
[...]
Now I would like to introduce our speaker for the session. Asya Bergal is a researcher at AI Impacts, where she also heads up their operations. She has a BA in computer science from MIT. Since graduating, she has worked as a trader and software engineer for Alameda Research, and as a research analyst at Open Philanthropy. Here's Asya.
**Asya:** Hi, my name is Asya Bergal. I work for an organization called AI Impacts, though the views in this presentation are mine and not necessarily my employer’s. I'm going to talk about some reasons why you might think [reaching] human-level AI soon is extremely unlikely.
I'll start with [some common] reasons. I'm interested in these because I'm interested in the question of whether we are, in fact, extremely unlikely to have human-level AI in the near term. [More specifically], is there a less-than-5% chance that we’ll see human-level AI in the next 20 years? I'm interested in this because if [reaching] human-level AI soon is extremely unlikely, and we know that now, that has some implications for what we, as effective altruists and people who care about the long-term future, might want to do.
If human-level AI is extremely unlikely, you might think that:
\* Broader movement-building might be more important, as opposed to targeting select individuals who can have impact now.
\* Fewer altruistic people who are technically oriented should be going into machine learning.
\* Approaches to AI safety should look more like foundational approaches.
\* We have more time for things like institutional reform.
\* There is some effect where, if the community is largely [broadcasting] that human-level AI [will soon occur], and it doesn't, we’ll lose some global trust in terms of having good epistemics and being right about things. Then, people will take us less seriously when AI risk is more of an imminent threat.
I don't know how real this last problem is, but it is something I worry about. Maybe it is a [potential] effect that we should be aware of as a community.
For the rest of this talk, I am going to look at three reasons I’ve heard that people [give to explain why] we won't have human-level AI soon:
1. A privileged class of experts disagree.
2. We're going to run out of “compute” [before we can reach] human-level AI.
3. Fuzzily defined current methods will be insufficient to get us there.
I don't claim that these three reasons are representative, but they're particularly interesting to me, and I spent some time investigating them. I will go ahead and spoil the talk now and say that I'm not going to answer the “5% question” [that I posed earlier], partially because my views on it vary wildly as I collect new evidence on some of these reasons. I do hope that I will, in the near term, get to the point where my views are pretty stable and I have something concrete to say.
**Reason No. 1: A privileged class of experts disagree**
Let’s look into the first reason — that experts might disagree that we can get to human-level AI soon. This is a survey conducted by Katja Grace from AI Impacts and a lot of other people. They asked machine learning researchers and experts what probability of human-level machine intelligence they think there will be [in a given] year.
You can see the 20-year mark [in the slide above]. It really seems like they do think there's more than a 5% chance.
Then, if you delve a little further into the survey results, you see that the answers people give are actually extremely sensitive to the framing and the exact way in which the question is asked — [for example,] whether you ask about human-level machine intelligence, automating all jobs, or the year in which there will be a 90% chance of something happening versus in another year.
For the question “What chance do we have of automating all jobs within 50 years?”, people give pretty low odds. What that tells me — and what others have concluded from this — is that it's very difficult to know what to do with surveys of experts. We probably shouldn't put a lot of weight on them.
I was particularly interested in a kind of survey where you ask people how much fractional progress they thought had been made toward human-level AI. You can naively extrapolate that to figure out how many years it will take until we get to 100%.
Robin Hanson did a version of this survey. He asked machine learning experts how far they had come in the last 20 years. All of the people he asked had worked in their sub-fields for at least 20 years. They answered in the 5-10% range, which, when naively extrapolated, puts human-level AI at 300 to 400 years away. That is pretty long.
Then, the Katja Grace survey that I mentioned earlier did a very similar thing, but that team surveyed people who had been working in the field for anywhere from two to many years. Their aggregated percentages forecasted human-level AI to be something like 36 years away — much shorter than Hanson's aggregated forecast.
Even if you set the condition of [only looking at responses from] experts working in the field for 20 years or more, they answer in the 20% or 30% range, which [results in] a median forecast of 140 years or so. That’s still a pretty long way off, but not as long as the Hanson survey.
There's a fairly consistent story by which you could reconcile these two results: In the last 20 years, there was a long period in which there wasn't much progress made in AI.
This is a very simplified graph showing that only recently has there been a lot of progress. Across these surveys, people consistently say that progress has been accelerating recently. If you naively extrapolate the past 20 years, you get a boom-bust pattern implying that we won’t have human-level AI for a long time — whereas if you somewhat naively extrapolate the past five years, and perhaps take into account the fact that things might be accelerating, you can get a “pretty soon” forecast.
It's not clear based on these survey results whether 20-year experts definitely think that we won’t have human-level AI soon.
**Reason No. 2: We’re going to run out of “compute” before we can reach human-level AI**
The second reason why you might think we won't soon attain human-level AI is that we're going to run out of “compute” [processing resources]. In an [analysis done by OpenAI](https://openai.com/blog/ai-and-compute/), researchers looked at the amount of compute used in the largest machine-learning experiments for training from 2012 to 2018. I also [included] [GPT-3](https://en.wikipedia.org/wiki/GPT-3), because OpenAI recently released [data on] how much compute they used to train it.
They noticed that over this period of time, there was a pretty consistent exponential rate of growth — around 11.5x per year. A natural question looking at this graph, and given that compute has historically been really important for machine-learning progress (and was important for these results), is: What will this trend look like in the future? Will it be faster? Will it be at the same rate? Will it be slower? What happens to AI progress as a result of this trend?
It's somewhat likely that it's going to slow down. Therefore, I'm particularly interested in asking, “What happens if it slows?” For the question “What will this trend look like in the future?”, it's reasonable to start by looking at what has happened with compute in the past. I've tried to illustrate that. I’ll now attempt to explain [the slide].
On the bottom, I put price performance (i.e. improvements in the amount of compute you can buy per dollar). On the top, I put general things that were happening in the world of machine learning training. For a long time, most computation and ML training was done on CPUs, which were governed by [Moore's law](https://en.wikipedia.org/wiki/Moore%27s_law) and showed a 1.4x per year increase in price performance. Then, around 2008, the price performance increase [fell back] and looked more like 1.1x a year. Then, several things started happening around 2012.
One big [change] was that people started training neural networks and machine learning techniques on GPUs, which Nvidia estimated as a 35x improvement for at least one task. Then, starting in 2012 [and continuing today], in 2020, two major things happened. The biggest is that people are willing to spend way more money buying compute, whereas from 2012 to 2016, you may have seen people training on one to eight GPUs. Closer to 2018, you can see that people are using hundreds of GPUs, and the techniques required to enable training on a lot of GPUs at once are also improving. Huge amounts of parallelization [became increasingly feasible].
[Meanwhile], a much smaller effect was that the price performance [the amount of processing power you can buy with a given amount of money] within an individual GPU or an individual piece of hardware improved. From 2012 to 2016, I estimate this to have been about 1.2x a year. Then, around 2016, several different companies started creating hardware that was especially optimized for deep learning. Between 2016 to 2020 I estimate that there was a 1.5x to 3x increase in price performance.
The main factor is that people are increasing their spending. A natural question to ask, then, if you want to know what this trend might look like in the future, is: How much more will people spend? How much more price performance will we see?
[Although spending has] powered growth recently, our spending can’t continue to increase for very long. People estimate that the 2018 experiment cost around $10 million. If we wanted to match the previous trend of 11.5x a year, we could only get to 2022 on spending alone. If we spent $200 billion, which is 1% of US GDP — the amount we [could] see if governments are really interested in machine learning — we would have to compensate largely with price performance after two years.
Looking at price performance and naively extrapolating this 3x [increase] from a period when people really started optimizing for this, the result is still that we won’t match the 11.5x that we saw before. [Transcriber’s note: actual yearly increases may be lower if most low-hanging fruit was taken during the initial optimization process, making further increases more difficult.]
I do think there are a lot of reasons that we should think it's plausible for machine learning to go faster in the coming years. Maybe companies like Nvidia and Google invest even more of their resources into AI, and make even more improvements. Maybe some of the startups coming out with specialized chips for AI do very well; maybe if you design a chip specifically for training on a particular type of deep-learning task, then you get much better performance.
Still, we should be aware of the fact that specialization gains have to end at some point, because you're not really making any new fundamental technological advances. You're just rearranging the bits in the computer to be very good at a particular thing. Eventually, we should expect that kind of specialization to end, and for AI to pull back to the trend in Moore's law that we saw before.
I think these two questions are important:
1. How quickly will improvements come?
2. How much [progress will] specialization improvements yield in total?
I think we must answer these in order to estimate what compute progress — and therefore AI progress — might look like in the future. That's the kind of work I'm interested in doing now.
My current best guess is that, given the impossibility [of spending continuing to increase] and the unlikeliness of price performance matching the gains we've seen in the past, we should expect the growth of compute to slow down. As I mentioned, it's very important to determine exactly how much. Then the question becomes: If this trend does slow down, then what happens?
Historically, compute has definitely been important, but perhaps we're now in a period where there's just a steady stream of investment. Perhaps we can compensate for compute that doesn't grow at the same exponential rate with better algorithms and efficiency gains.
On this particular question, I'm excited about two papers coming out of [Neil Thompson's lab at MIT](http://www.neil-t.com/moores-law-and-computer-performance/). One of them is called “How far can deep learning take us? A look at performance and economic limits.” The other is “The importance of exponentially more computing power,” which will look at the effects of computing power across several important domains. I’m plugging these papers because you should be excited about them, too.
In general, on this compute question, the future is still up in the air. For me, it's the defining question — and the most tractable one right now — in terms of what future progress to expect. I also think what comes out of big companies and startups in the next five years will be a good metric for thinking about AI progress over the next 20 years. I'm really excited to see what happens.
**Reason No. 3: Current methods are insufficient**
The third reason why you might think we're very unlikely to reach human-level AI in the near term is that current methods (which could refer to deep learning or neural nets in general) will be insufficient. I want to split [this idea] into two categories.
One is that the methods will be fundamentally insufficient somehow. Several legitimate computer scientists have reasons for thinking that, given how current techniques look, we're not going to get to a human level. One reason is that human intelligence relies on many priors about the world, and there’s no clear way to [inject] those priors into neural-network architectures.
Another reason is that we might not think that neural networks are going to be able to build the kind of causal models that human reasoning relies on. Maybe we think that neural networks and deep learning systems won’t be able to deal with hierarchical structure. Maybe we think that we won’t be able to collect all of the data that we would need to train something that's actually human-level.
I don't think I have the technical expertise to evaluate these [theories]. But in my non-expert opinion, we don’t have enough evidence about these methods to be able to say that they're fundamentally insufficient in any of these ways. As with compute, I think in the next five years a lot of [research will likely be released] that sheds light on what we can expect from these methods over the next 20.
I view the other broad class of “insufficiency reasons” as the argument that current methods might be insufficient from a practicality standpoint — that we could get to human-level AI with neural nets and deep learning methods, but we won't because it'll just be too difficult. Maybe we're in a bubble of hype and investment. Given that AI progress has to be powered by investment, we need a steady stream of money flowing in, which means a steady stream of economic value encouraging people to spend more money.
People often give the example of self-driving cars as an advance that we thought [would become mainstream] years ago, but has taken much longer than we anticipated. Maybe you can generalize from there, and claim that many human tasks are very difficult to automate — and that it will be a long time before we can get value out of automating them. You might think that investment will dry up once we successfully automate the small set of human tasks that are easy for neural networks to automate.
Another general argument that people use is the idea that in scientific fields we should expect diminishing returns. We should expect fewer good things [to emerge from] neural networks and deep learning over time. It's hard to use this argument, partially because it's not clear when we should expect those diminishing returns to kick in. Maybe we should expect that to happen after we get to human-level AI. Although it has historically seemed somewhat true that we should expect diminishing returns in a lot of fields.
The last crux I want to point to [of the “practicality” argument] relates to [what people believe about] the amount of work that we need before we can create a system that's fairly general. One model of AI progress suggests that we're slowly automating away human jobs, one job at a time. If you think we would need to automate away all jobs before [AI can reach human-level intelligence], then you might think it will take a long time, especially given that we haven't yet automated even the most basic jobs.
There's another model proposing that generality isn't all that far away. This model indicates that once we have something that is essentially general, it will be able to automate a lot of the jobs away. In that case, you're looking for something that approximates the human brain — and you might think that will happen soon. This question of when we arrive at something general, and whether that happens before or after automating everything else, is in people's minds when they disagree about whether [human-level AI is practical or not].
In conclusion, do experts disagree that we could have human-level AI soon? That’s not obvious to me. Even if you only consider the opinions of experts who have worked in the field for a long time, I don't think it's clear that they disagree.
Will we run out of compute? I’m still working on this one. My current guess is that we won't maintain our current growth rates for the next 20 years, but I'm not sure that we should expect that to cause progress to slow significantly. That's a harder question to answer.
Then, are current methods insufficient? We don't have evidence now that strongly suggests an answer one way or the other. But again, I do expect a lot of work to emerge in the coming years that I think will shed light on some of these questions. We might have better answers pretty soon.
That was my talk. I hope you enjoyed it. If you have questions, think I was wrong about something, or want clarification on anything, please feel free to [reach out to me] — especially if your question doesn't get answered in the upcoming Q and A. I try to be very approachable and responsive.
**Aaron:** Thank you for that talk, Asya. We’ve had a number of questions submitted already, so let's kick off the Q and A with the first one: What have you changed your mind about recently?
**Asya:** I only recently looked carefully at the data I shared in this talk on compute trends. Before that, I was thinking, “Well, who knows how fast these things are improving? Maybe we will just compensate by increasing price performance a lot.”
Seeing the new data — and that, at least so far, a lot of these hardware startups haven't been that successful — made me feel a little more skeptical about the future hardware situation. Again, that's very tentative. I'll change my mind again next week.
**Aaron:** Yes, and who knows what else quantum computing or such things could bring about?
If you had six or more months for further research, what would your priorities be — or what might someone else interested in this topic [focus on]?
**Asya:** Hopefully, I will do future research. I’m quite interested in the compute question, and there are several things you can do to try to estimate the effects. You can talk to people and look at what specific improvements have led to in other fields.
There's a lot of economic data on how much you have to spend on specialized hardware to see efficiency improvements. Looking at historical data on specializations like Bitcoin could provide some idea of what we can expect hardware specialization to look like for deep learning. Again, this is very new. We'll see what I end up looking into.
**Aaron:** Sounds cool. How useful do you think economic growth literature is for forecasting AGI [artificial general intelligence] development, in terms of timelines?
**Asya:** I've been looking into this recently. The macroeconomics work that I've seen [comprises] a lot of abstract models with different variables and guesses as to how AI will affect automation, how automation will affect several other parameters, and how those will affect economic growth. They're usually very theoretical and involve guessing at a broad swath of [possibilities].
This might be very difficult, but I'd be interested in more empirical work — specifically, the particulars of automation in the supply chain in AI. What would we expect to be automated? At what rate would we expect jobs to be automated? One common empirical model is based on automating away a constant fraction of jobs every year, or something to that effect. I don't know if that's a reasonable assumption.
If we had more empirical data, I'd be pretty excited about economic growth modeling. That requires work beyond the growth models that exist right now.
**Aaron:** Sounds perfect. Will the future of compute also be limited by pure physics? Faster chips are now seven nanometers [one billionth of a meter], but we can't shrink that indefinitely. Will that potentially limit the growth toward AGI?
**Asya:** There are definitely pure, Moore’s-law physical limits, especially once we get past specialization. You mentioned a wide swath of more exotic architectures that we can exploit before the physical limits become the most relevant factor — things like quantum, optical, stacking transistors in 3-D space, etc. I think we can [leverage] improvements from those before we have to worry about the fact that we are hitting physical limits in transistor density.
But I do think it's relevant. We shouldn't expect progress at the same rates that we used to see, because we really just can't do that anymore.
**Aaron:** Totally makes sense. Do you think that figuring out when we can expect discontinuous progress in AI can tell us much about whether AGI will happen soon?
**Asya:** It does somewhat, especially if you have a concrete metric in mind and if you don't think progress will be discontinuous, which the AI Impacts investigation suggests is unlikely (but it could be likely). Then, if you want to see continuous progress that leads to AGI in the next 20 years, you have to expect signs of fairly steep, exponential growth.
If you're not seeing those signs, and you think progress is likely to be continuous (i.e. there's not going to be a huge jump), you shouldn’t expect AGI to come soon.
So, I think it should influence people's opinions, especially if we expect something like the economic value from AI or AGI to increase continuously, and it's not increasing very much. In that case, maybe we can say something about what the future looks like.
**Aaron:** Sounds good. Even if AGI isn't near, shouldn't we worry about the side effects of increasingly powerful machine learning shaping our world with misaligned incentives?
**Asya:** We totally should. When I’m thinking about this, I consider the more nebulous risks that you get into with superhuman agents. A subset of the literature on AI risk concerns is very focused on those. In some sense, they are some of the most neglected — and the most likely to be existentially bad.
I definitely think powerful machine learning systems of various kinds that are not at a human level can be transformative and quite impactful. It's societally worth thinking about that for sure.
**Aaron:** Perfect. What do you think about AI talent as another factor in AI development? Will the decoupling between the US and China slow down AI research?
**Asya:** I should qualify my answer: I'm probably not the most qualified person to talk about AI talent in the US and China. I'm not sure if we should expect it to slow down research, because I'm not sure that it was particularly coupled before in a way that was essential for research. Therefore, I'm not sure that the decoupling should imply a slowdown, but again, I'm definitely not an expert in either AI talent or China. Don't take my opinions on this too seriously.
**Aaron:** Makes sense. Are there any specific, “canary in the coal mine” advances that we can [use to determine] that human-level AI will happen within 10 years?
**Asya:** I don't know — perhaps if there was good evidence that a lot of the theoretical problems that people think are true of deep learning systems were shown to be untrue (for example, if we suddenly felt like we could create really good simulation environments, so that training wasn't a problem). Anything that knocks out problems that are [impossible to chart with] causal modeling or hierarchical structures because they need too much data would make me much more optimistic about deep learning methods getting us to AGI in 20 years.
**Aaron:** Makes sense. Do you think using forecasting methods similar to the ones you've used could help predict when another AI winter could potentially happen?
**Asya:** I'm not really sure that I would say “similar to the ones I've used.” The “AI winter” part is a function of investment. Right now, investment looks to be exponentially increasing. Post-coronavirus, it would be interesting to see what investment looks like. That will be the type of thing that I'll be keeping track of in terms of predicting an AI winter: investment levels and how many researchers are entering machine learning PhD programs.
**Aaron:** Makes sense. What do you think the policy implications are for different timelines of AI progress?
**Asya:** That's a really good question. I think it's probable that if AI becomes a strategically important technology, there will be some government involvement. Then there's a question of what the government will actually do. One function of the government is to legislate the things that are inputs to AI: AI labs, compute manufacturers in the US, etc.
If timelines are short, given that the policy world moves rather slowly, it's important to be on the ground and figuring that out now. but if timelines are long, we have more time for things like institutional reform, and don't need to be in as much of a rush getting laws, cooperation mechanisms, and things like that implemented now.
**Aaron:** Cool. Expert surveys like the one you cited in your talk are often used as a starting place, such as in Toby Ord's [*The Precipice*](https://www.amazon.com/dp/B07VB299G3/). Do you think there should be a revised version of the survey every year? Is it an appropriate starting point for researchers such as yourself to use these estimates?
**Asya:** It's not unreasonable if you don't know something about a field to ask experts what they think. From that perspective, I don't know that we have a better starting point for random forecasting exercises.
Should such a survey be done every year? It would be cool if it were. It would certainly be interesting to know how expert opinion is changing. But I don't know that we should if better forecasting tools [emerge]. If we are able to analyze economic trends and things like compute, I’m more inclined to endorse that as a starting point. As I mentioned in the talk, I've become pretty skeptical of surveys of experts. I'm not sure that a random technical expert actually has a more informed view than we do.
I do think it's an easy way to determine a good starting point. I think I do endorse frequent surveys of this type. If nothing else, it's interesting to see technical researchers’ perceptions of progress; that's a valuable data point in and of itself, even if it doesn't tell us anything about our true progress.
**Aaron:** Yes. Is it possible that current spending on AI as a percentage of GDP will stay constant, or even diminish, because of the huge gains in GDP we might expect from investment in AI?
**Asya:** Over the long term that's very possible. I don't expect it in the short term. But you could imagine in a world where AI contributes to huge economic growth — where we spend more on AI, but GDP increases because we're actually spending less in terms of a percentage. I do think that's possible, but I would find it pretty unlikely in the short term.
**Aaron:** Do timescales change the effect of interventions for suffering-focused or s-risks in your opinion?
**Asya:** That's another good question that I haven't thought about at all. I don't know. Wow, this is going to reveal how little I've thought about s-risks. At least some of the work on s-risks is related to work on AGI in that we don't want astronomical amounts of suffering to be created. In the same way that we are, even without S-risks, worried about risks from AGI, there are similar implications to the timescales on AGI that are related to s-risks. I'm not very well-read on the s-risk space.
**Aaron:** Perfect. I think we have time for one last question: Given the uncertainty around AGI and timelines, would you advise donors to hold off on assisting AI safety organizations?
**Asya:** I think not. If work seems as though it’s plausibly good, it’s [worthwhile to donate]. I think direct work is often instrumentally useful for [guiding] future direct work and for things like movement-building. Therefore, I would not hold off on funding safety work now — especially safety work that seems good or promising.
**Aaron:** Thank you so much, Asya. That concludes the Q and A part of the session. |
1bb70040-0208-40b4-b938-a9da214bd7cc | trentmkelly/LessWrong-43k | LessWrong | Meetup : Reading First Meetup!
Discussion article for the meetup : Reading First Meetup!
WHEN: 07 October 2015 06:00:00PM (+0100)
WHERE: Reading, England
We'll be meeting in the Starbucks next to Vue cinema and the river Kennet, in the Oracle Shopping Centre (Riverside Entrance). I'll be there from 6pm and I'll stay for at least two hours (longer, if enough people turn up to make it a proper meeting). There will be at least one other person there (so, at least two of us). I will have a whiteboard with paperclips drawn on it and "LessWrong" written on it, wear my EAG shirt and have a makeshift Slytherin badge on (there is at least an 85% chance I forget at least one of those things, in which case just look for the group of geeks - it's quite a small Starbucks so it ought to be obvious enough).
This is our first meeting and we're planning to make it a regular thing if we get enough interest, so this one will probably be focused on introducing ourselves and discussing plans. We'll hopefully announce more, so don't worry if you miss the first one.
https://www.google.co.uk/maps/place/The+Oracle+Shopping+Centre/@51.45308,-0.971488,3a,75y,90t/data=!3m8!1e2!3m6!1s7305926!2e1!3e10!6s%2F%2Fstorage.googleapis.com%2Fstatic.panoramio.com%2Fphotos%2Fsmall%2F7305926.jpg!7i2048!8i1536!4m2!3m1!1s0x48769b15e736aff1:0x91a4b0447e720259!6m1!1e1 <--- Here is a photo of the area to be in, in case the location directions weren't clear enough. People in IRC say the link works. You have to kind of squint to see the Starbucks in the photo, but if you recognise this, you're close.
Hope to see you there :)
Discussion article for the meetup : Reading First Meetup! |
af1bd89c-cc08-4eba-ac7e-13ea30c77455 | trentmkelly/LessWrong-43k | LessWrong | on Science Beakers and DDT
tech trees
There's a series of strategy games called Civilization. In those games, the player controls a country which grows and develops over thousands of years, and science is one of the main types of progress. It involves building facilities to generate research points, sometimes represented by Science Beakers filled with Science Fluid, and using those points to buy tech nodes on a tree.
I'm reminded of the above game mechanic when I see certain comments on blogs or Twitter/X. I'll use supersonic aircraft as an example.
supersonic aircraft
Several times, I've seen comments online similar to: "Look at this graph of aircraft speeds going up and then stopping. People should be forced to deal with sonic booms so we can have supersonic aircraft and get back to progress. It will be unpopular but this is more important than democracy."
In their mind, society has a "tech node" called "supersonic flight" and you need to finish the node before you can move to the next node. If opponents cause society to not finish the node, then you're stuck. But that's not how tech progress works. How much further would we be from supersonic transports today if the Concorde had never flown? The answer is 0.
There's a lot of underlying progress required for something like, say, a high-performance gas turbine. The abstract progress is more general, and each practical application involves different specifics. Understanding principles of metallurgy and how to examine metals leads to specific high-temperature alloys, which are what actually get used. Funding for gas turbine development can lead to single-crystal casting or turbine blade film cooling being developed, but the actual production of a new gas turbine model has no effect on whether that's developed.
The tech tree of a game like Civilization would be more accurate if it had a 3rd dimension, with height representing the abstract and general vs the practical and specific. Something like the Concorde would be on a vertical offshoo |
b5649fa2-40e3-4ff6-85c1-89ea5bfe30e0 | trentmkelly/LessWrong-43k | LessWrong | How do you optimize productivity with respect to your menstrual cycle?
As the question asks, I am looking for productivity advice directly or indirectly related to one's menstrual cycle.
Motivation: This may seem like a gross or weirdly personal question but I think it's actually quite important. For example, for the first four hours of my period, I am unable to do anything besides lie down, be in pain, and (sometimes) vomit for about 4 hours. Not accounting for all of the other ways my period negatively affects me, this is already 4 hours * 12 = 48 hours = 2 full days lost per year. That being said, there can be menstruation related hormonal fluctuations which positively influence productivity. During ovulation, I can sometimes reach almost manic levels of mental and social energy and (I'm less sure about this) I require a bit less sleep to function.
What the question is asking: This is a very open-ended and personal question so I am expecting similarly open-ended and personal answers. For example, stuff I'm interested in:
* Generally speaking, how does your menstrual cycle affect your productivity?
* What effects has the pill had? Which kind of pill?
* Have you experimented with varying your cycle length? (I had a 3 month cycle some number of years ago, and enjoyed the decreased frequency and intensity of my periods.)
* Have you noticed fluctuations in energy level caused by your menstrual cycle? Any way to make the high energy phases longer and the low energy phases shorter? Do phases of high energy correspond to a certain hormone which I could get through the pill?
* Are there dietary changes that mitigate the bad effects of menstruation / enhance the good effects? (For example, is it "helpful" to give in to cravings?)
* Are there sleep schedule choices that can be made which synergize with your menstrual cycle?
* Are there points in your cycle where you are best at detail-oriented work / creative work / socializing / writing / anything else relevant?
It's possible that your answers to many of these questions are "my |
bd18aec8-b9a6-4445-a5b5-a448bc2bf63d | StampyAI/alignment-research-dataset/arxiv | Arxiv | Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations
1 Introduction
---------------
Deep neural networks have been very successful across a variety of tasks.
This success has also driven applications in domains where reliability and security are critical, including self-driving cars (Bojarski et al., [2016](#bib.bib7)), health care, face recognition (Sharif et al., [2017](#bib.bib25)), and the detection of malware (LeCun et al., [2015](#bib.bib17)).
Security concerns emerge when an agent using the system could benefit from the system performing poorly.
Reliability concerns come about when the distribution of input data seen during training can differ from the distribution on which the model is evaluated.
*Adversarial examples* (Goodfellow et al., [2014](#bib.bib12)) is a method to attack neural network models. This attack applies a small perturbation to the input that changes the predicted class. It is notable that it is possible to produce a perturbation small enough that it is not noticeable with a naked eye.
It has been shown that simple gradient methods allow one to find a modification of the input that often changes the output class (Szegedy et al., [2013](#bib.bib26); Goodfellow et al., [2014](#bib.bib12)). More recent work demonstrated that it is possible to create a patch such that even when presented on the camera, it changes the output class with high confidence (Brown et al., [2017](#bib.bib8)).
Defences against adversarial examples have been developed as a response. Some of the most prominent classes of defences include feature squeezing (Xu et al., [2017](#bib.bib28)), adapted encoding of the input (Jacob Buckman, [2018](#bib.bib15)), and distillation-related approaches (Papernot et al., [2015](#bib.bib21)).
Existing defenses provide some robustness but most are not easy to deploy. In addition, many have been shown to be vulnerable to gradient masking. Still others require training a generative model directly in the visible space, which is still difficult today even on relatively simple datasets.
Our goal is to provide a method which
(i) can be generically added into an existing network;
(ii) robustifies the network against adversarial attacks and
(iii) provides a reliable signal of the existence of input data that do not lie on the manifold on which it the network trained.
The ability of generative models, used directly on the input data, to improve robustness is not new.
Our main contribution is that we employ this robustification on the distribution on the learned hidden representations instead making the identification of off-manifold examples easier Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations").
We propose Fortified Networks \symrook.
Fortification consists of using denoising autoencoders to “decorate” the hidden layers of the original network. We use “decoration” in the Pythonic sense that it can be applied to any function (in this case a part of the network) and extend its behavior without explicitly modifying it.
Fortification meets all three goals stated above.
We discuss the intuition behind the fortification of hidden layers and lay out some of the method’s salient properties.
We evaluate our proposed approach on MNIST, Fashion-MNIST, CIFAR10 datasets against whitebox and blackbox attacks.
Figure 1: Illustration of the autoencoder dynamics in the input space (top) and in abstract hidden space (bottom). The leftmost panels show data points from three different classes, the middle panels show vector fields describing the autoencoder dynamics, and the rightmost panels show a number of resulting trajectories and basins of attraction. The key motivation behind Fortified Networks is that directions which point off the data manifold are easier to identify in an abstract space with simpler statistical structure, making it easier to map adversarial examples back to the projected data manifold.
The rest of the paper is structured in the following way. Section [2](#S2 "2 Background ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations") gives a detailed overview of the background on the adversarial attacks and denoising autoencoders used in this work. Section [3](#S3 "3 Fortified Networks ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations") presents our proposed methods for the defence against adversarial examples, Section [4](#S4 "4 Experiments ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations") describes the experimental procedure and Section [5](#S5 "5 Results ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations") provides the experimental results and a comparison to previous approaches. Finally, Section [6](#S6 "6 Related Work ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations") puts this work into the context of previous publications and Section [7](#S7 "7 Conclusion ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations") concludes.
2 Background
-------------
###
2.1 The Empirical Risk Minimization Framework
Let us
consider a standard classification task with an underlying data distribution 𝒟𝒟\mathcal{D}caligraphic\_D
over pairs of examples x∈ℝd𝑥superscriptℝ𝑑x\in\mathbb{R}^{d}italic\_x ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_d end\_POSTSUPERSCRIPT and corresponding labels y∈[k]𝑦delimited-[]𝑘y\in[k]italic\_y ∈ [ italic\_k ]. We also
assume that we are given a suitable loss function L(θ,x,y)𝐿𝜃𝑥𝑦L(\theta,x,y)italic\_L ( italic\_θ , italic\_x , italic\_y ), for instance the
cross-entropy loss for a neural network. As usual, θ∈ℝp𝜃superscriptℝ𝑝\theta\in\mathbb{R}^{p}italic\_θ ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_p end\_POSTSUPERSCRIPT is the set of
model parameters. Our goal then is to find model parameters θ𝜃\thetaitalic\_θ that minimize
the risk 𝔼(x,y)∼𝒟[L(x,y,θ)]subscript𝔼similar-to𝑥𝑦𝒟delimited-[]𝐿𝑥𝑦𝜃\mathbb{E}\_{(x,y)\sim\mathcal{D}}[L(x,y,\theta)]blackboard\_E start\_POSTSUBSCRIPT ( italic\_x , italic\_y ) ∼ caligraphic\_D end\_POSTSUBSCRIPT [ italic\_L ( italic\_x , italic\_y , italic\_θ ) ]. This expectation cannot be computed, therefore a common approach is to to minimize the empirical risk 1/N∑DL(x,y,θ)1𝑁subscript𝐷𝐿𝑥𝑦𝜃1/N\sum\_{D}L(x,y,\theta)1 / italic\_N ∑ start\_POSTSUBSCRIPT italic\_D end\_POSTSUBSCRIPT italic\_L ( italic\_x , italic\_y , italic\_θ ) taking into account only the examples in a given dataset D𝐷Ditalic\_D.
###
2.2 Adversarial Attacks and Robustness
While the empirical risk minimization framework has been very successful and often leads to excellent generalization, it has the significant limitation that it doesn’t guarantee robustness, and more specifically performance on examples off the data manifold. Madry et al. ([2017](#bib.bib19)) proposed an optimization view of adversarial robustness, in which the adversarial robustness of a model is defined as a min-max problem
| | | | | |
| --- | --- | --- | --- | --- |
| | minθρ(θ)subscript𝜃𝜌𝜃\displaystyle\min\_{\theta}\rho(\theta)roman\_min start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT italic\_ρ ( italic\_θ ) | , where \displaystyle,\quad\text{ where }\quad, where | | (1) |
| | ρ(θ)𝜌𝜃\displaystyle\rho(\theta)italic\_ρ ( italic\_θ ) | =𝔼(x,y)∼𝒟[maxδ∈𝒮L(θ,x+δ,y)].absentsubscript𝔼similar-to𝑥𝑦𝒟delimited-[]subscript𝛿𝒮𝐿𝜃𝑥𝛿𝑦\displaystyle=\mathbb{E}\_{(x,y)\sim\mathcal{D}}\left[\max\_{\delta\in\mathcal{S}}L(\theta,x+\delta,y)\right]\;.= blackboard\_E start\_POSTSUBSCRIPT ( italic\_x , italic\_y ) ∼ caligraphic\_D end\_POSTSUBSCRIPT [ roman\_max start\_POSTSUBSCRIPT italic\_δ ∈ caligraphic\_S end\_POSTSUBSCRIPT italic\_L ( italic\_θ , italic\_x + italic\_δ , italic\_y ) ] . | | (2) |
###
2.3 Denoising Autoencoders
*Denoising autoencoders* (DAEs) are neural networks which take a noisy version of an input (for example, an image) and are trained to predict the noiseless version of that input. This approach has been widely used for feature learning and generative modeling in deep learning (Bengio et al., [2013a](#bib.bib5)). More formally, denoising autoencoders are trained to minimize a reconstruction error or negative log-likelihood of generating the clean input. For example, with Gaussian log-likelihood of the clean input given the corrupted input, r𝑟ritalic\_r the learned denoising function, C𝐶Citalic\_C a corruption function with Gaussian noise of variance σ2superscript𝜎2\sigma^{2}italic\_σ start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT, the reconstruction loss is
| | | | |
| --- | --- | --- | --- |
| | ℒ^=1N∑n=1N(‖r(Cσ(x(n)))−x(n)‖22).^ℒ1𝑁superscriptsubscript𝑛1𝑁superscriptsubscriptnorm𝑟subscript𝐶𝜎superscript𝑥𝑛superscript𝑥𝑛22\widehat{\mathcal{L}}=\frac{1}{N}\sum\_{n=1}^{N}\left(\left\|r(C\_{\sigma}(x^{(n)}))-x^{(n)}\right\|\_{2}^{2}\right).over^ start\_ARG caligraphic\_L end\_ARG = divide start\_ARG 1 end\_ARG start\_ARG italic\_N end\_ARG ∑ start\_POSTSUBSCRIPT italic\_n = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_N end\_POSTSUPERSCRIPT ( ∥ italic\_r ( italic\_C start\_POSTSUBSCRIPT italic\_σ end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUPERSCRIPT ( italic\_n ) end\_POSTSUPERSCRIPT ) ) - italic\_x start\_POSTSUPERSCRIPT ( italic\_n ) end\_POSTSUPERSCRIPT ∥ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT ) . | | (3) |
Alain et al. ([2012](#bib.bib1)) demonstrated that with this loss function, an optimally trained denoising autoencoder’s reconstruction vector is proportional to the gradient of the log-density:
| | | | |
| --- | --- | --- | --- |
| | rσ(x)−xσ2→∂logp(x)∂xasσ→0.formulae-sequence→subscript𝑟𝜎𝑥𝑥superscript𝜎2𝑝𝑥𝑥as→𝜎0\frac{r\_{\sigma}(x)-x}{{\sigma^{2}}}\rightarrow\frac{\partial\log p(x)}{\partial x}\hskip 10.00002pt\textrm{as}\hskip 10.00002pt{\sigma}\rightarrow 0.divide start\_ARG italic\_r start\_POSTSUBSCRIPT italic\_σ end\_POSTSUBSCRIPT ( italic\_x ) - italic\_x end\_ARG start\_ARG italic\_σ start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT end\_ARG → divide start\_ARG ∂ roman\_log italic\_p ( italic\_x ) end\_ARG start\_ARG ∂ italic\_x end\_ARG as italic\_σ → 0 . | | (4) |
This theory establishes that the reconstruction vectors from a well-trained denoising autoencoder form a vector field which points in the direction of the data manifold. However, Alain et al. ([2012](#bib.bib1)) showed that this may not hold for points which are distant from the manifold, as these points are rarely sampled during training. In practice, denoising autoencoders are not just trained with tiny noise but also with large noises, which blurs the data distribution as seen by the learner but makes the network learn a useful vector field even far from the data.
3 Fortified Networks
---------------------
Figure 2: An illustration of the process of mapping back to the manifold in the visible space (left) and the hidden space (right). The shaded regions represent the areas in the space which are occupied by data points from a given class (they do *not* represent decision boundaries).
pty𝑦yitalic\_yy~~𝑦\widetilde{y}over~ start\_ARG italic\_y end\_ARGpthkdecodedsuperscriptsubscriptℎ𝑘𝑑𝑒𝑐𝑜𝑑𝑒𝑑h\_{k}^{decoded}italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_d italic\_e italic\_c italic\_o italic\_d italic\_e italic\_d end\_POSTSUPERSCRIPThk~decodedsuperscript~subscriptℎ𝑘𝑑𝑒𝑐𝑜𝑑𝑒𝑑\widetilde{h\_{k}}^{decoded}over~ start\_ARG italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG start\_POSTSUPERSCRIPT italic\_d italic\_e italic\_c italic\_o italic\_d italic\_e italic\_d end\_POSTSUPERSCRIPTptEncoded𝐸𝑛𝑐𝑜𝑑𝑒𝑑Encodeditalic\_E italic\_n italic\_c italic\_o italic\_d italic\_e italic\_dEncoded𝐸𝑛𝑐𝑜𝑑𝑒𝑑Encodeditalic\_E italic\_n italic\_c italic\_o italic\_d italic\_e italic\_dptInject NoiseInject Noisepthksubscriptℎ𝑘h\_{k}italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPThk~~subscriptℎ𝑘\widetilde{h\_{k}}over~ start\_ARG italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARGptx𝑥xitalic\_xx~~𝑥\widetilde{x}over~ start\_ARG italic\_x end\_ARGDec𝐷𝑒𝑐Decitalic\_D italic\_e italic\_cEnc𝐸𝑛𝑐Encitalic\_E italic\_n italic\_cDec𝐷𝑒𝑐Decitalic\_D italic\_e italic\_cEnc𝐸𝑛𝑐Encitalic\_E italic\_n italic\_cℒadvsubscriptℒ𝑎𝑑𝑣\mathcal{L}\_{adv}caligraphic\_L start\_POSTSUBSCRIPT italic\_a italic\_d italic\_v end\_POSTSUBSCRIPTℒrecsubscriptℒ𝑟𝑒𝑐\mathcal{L}\_{rec}caligraphic\_L start\_POSTSUBSCRIPT italic\_r italic\_e italic\_c end\_POSTSUBSCRIPTℒc(y)subscriptℒ𝑐𝑦\mathcal{L}\_{c}(y)caligraphic\_L start\_POSTSUBSCRIPT italic\_c end\_POSTSUBSCRIPT ( italic\_y )ℒc(y~)subscriptℒ𝑐~𝑦\mathcal{L}\_{c}(\widetilde{y})caligraphic\_L start\_POSTSUBSCRIPT italic\_c end\_POSTSUBSCRIPT ( over~ start\_ARG italic\_y end\_ARG )Fortified Layer\symrook
Figure 3: Diagram illustrating a one-layer fortified network. A network is evaluated with a data sample x𝑥xitalic\_x and its corresponding adversarial example x~~𝑥\widetilde{x}over~ start\_ARG italic\_x end\_ARG. Hidden units hksubscriptℎ𝑘h\_{k}italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT and hk~~subscriptℎ𝑘\widetilde{h\_{k}}over~ start\_ARG italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_ARG are corrupted with noise, encoded with the encoder Enc𝐸𝑛𝑐Encitalic\_E italic\_n italic\_c, and decoded with the decoder Dec𝐷𝑒𝑐Decitalic\_D italic\_e italic\_c. The autoencoder (denoted by the red color) is trained to reconstruct the hidden unit hksubscriptℎ𝑘h\_{k}italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT that corresponds to the clean input. Dotted lines are two reconstruction costs: for a benign (ℒrecsubscriptℒ𝑟𝑒𝑐\mathcal{L}\_{rec}caligraphic\_L start\_POSTSUBSCRIPT italic\_r italic\_e italic\_c end\_POSTSUBSCRIPT) and adversarial examples (ℒadvsubscriptℒ𝑎𝑑𝑣\mathcal{L}\_{adv}caligraphic\_L start\_POSTSUBSCRIPT italic\_a italic\_d italic\_v end\_POSTSUBSCRIPT). Note that a layer can be fortified at any position in the network.
We propose the use of DAEs inserted at crucial points between layers of the original neural network in order to clean up the transformed data points which may lie away from the original data manifold.
Intuitively, the method aims to regularize the hidden representations by keeping the activations on the surface of the corresponding projected data manifold through the application of a DAE trained on the hidden representations (on the original clean data).
We argue that applying the DAEs on the hidden layers—as opposed to the raw input signal—facilitates learning, while providing a stronger protection from adversarial attacks. As illustrated in Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations"), we hypothesize that more abstract representations associated with deeper networks are easier to clean up because the
transformed data manifolds are flatter. The flattening of data manifolds in the deeper layers of a neural network was first noted experimentally by Bengio et al. ([2013b](#bib.bib6)).
We provide experimental support for these claims in Section [4](#S4 "4 Experiments ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations").
#####
\symrook Layer fortification \symrook
Our method works by substituting a hidden layer hksubscriptℎ𝑘h\_{k}italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT with a denoised version.
We feed the signal hksubscriptℎ𝑘h\_{k}italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT through the encoder network, Eksubscript𝐸𝑘E\_{k}italic\_E start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT, and decoder network, Dksubscript𝐷𝑘D\_{k}italic\_D start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT, of a DAE for layer k𝑘kitalic\_k, which yields the denoised version, hkdecodedsuperscriptsubscriptℎ𝑘𝑑𝑒𝑐𝑜𝑑𝑒𝑑h\_{k}^{decoded}italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_d italic\_e italic\_c italic\_o italic\_d italic\_e italic\_d end\_POSTSUPERSCRIPT:
| | | | |
| --- | --- | --- | --- |
| | hkdecoded=Dk(Ek(hk+nk)),superscriptsubscriptℎ𝑘𝑑𝑒𝑐𝑜𝑑𝑒𝑑subscript𝐷𝑘subscript𝐸𝑘subscriptℎ𝑘subscript𝑛𝑘h\_{k}^{decoded}=D\_{k}(E\_{k}(h\_{k}+n\_{k})),italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_d italic\_e italic\_c italic\_o italic\_d italic\_e italic\_d end\_POSTSUPERSCRIPT = italic\_D start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ( italic\_E start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ( italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT + italic\_n start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ) ) , | | (5) |
where nksubscript𝑛𝑘n\_{k}italic\_n start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT is white Gaussian noise of variance σ2superscript𝜎2\sigma^{2}italic\_σ start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT and appropriate shape.
We call the resulting layer, a fortified layer and the resulting network the fortified network corresponding to the original network.
For training purposes, we treat the DAEs as part of the fortified network, backpropagate through and train all weights jointly.
Aside from the original classification loss, ℒcsubscriptℒ𝑐\mathcal{L}\_{c}caligraphic\_L start\_POSTSUBSCRIPT italic\_c end\_POSTSUBSCRIPT,
we also include the classification loss from the adversarial objective, ℒc(y~)subscriptℒ𝑐~𝑦\mathcal{L}\_{c}(\widetilde{y})caligraphic\_L start\_POSTSUBSCRIPT italic\_c end\_POSTSUBSCRIPT ( over~ start\_ARG italic\_y end\_ARG ) and
we introduce a dual objective for the DAEs.
* •
Reconstruction loss. For a mini-batch of N𝑁Nitalic\_N clean examples, x(1),…,x(N)
superscript𝑥1…superscript𝑥𝑁x^{(1)},\ldots,x^{(N)}italic\_x start\_POSTSUPERSCRIPT ( 1 ) end\_POSTSUPERSCRIPT , … , italic\_x start\_POSTSUPERSCRIPT ( italic\_N ) end\_POSTSUPERSCRIPT,
each hidden layer hk(1),…,hk(N)
superscriptsubscriptℎ𝑘1…superscriptsubscriptℎ𝑘𝑁h\_{k}^{(1)},\ldots,h\_{k}^{(N)}italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 1 ) end\_POSTSUPERSCRIPT , … , italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_N ) end\_POSTSUPERSCRIPT is fed into a DAE loss, similar to ([3](#S2.E3 "3 ‣ 2.3 Denoising Autoencoders ‣ 2 Background ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations")):
| | | |
| --- | --- | --- |
| | ℒrec,k=1N∑n=1N‖Dk(Ek(hk(n)+nk))−hk(n)‖22.subscriptℒ
𝑟𝑒𝑐𝑘1𝑁superscriptsubscript𝑛1𝑁superscriptsubscriptnormsubscript𝐷𝑘subscript𝐸𝑘superscriptsubscriptℎ𝑘𝑛subscript𝑛𝑘superscriptsubscriptℎ𝑘𝑛22\mathcal{L}\_{rec,k}=\frac{1}{N}\sum\_{n=1}^{N}\left\|D\_{k}\left(E\_{k}\left(h\_{k}^{(n)}+n\_{k}\right)\right)-h\_{k}^{(n)}\right\|\_{2}^{2}.caligraphic\_L start\_POSTSUBSCRIPT italic\_r italic\_e italic\_c , italic\_k end\_POSTSUBSCRIPT = divide start\_ARG 1 end\_ARG start\_ARG italic\_N end\_ARG ∑ start\_POSTSUBSCRIPT italic\_n = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_N end\_POSTSUPERSCRIPT ∥ italic\_D start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ( italic\_E start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ( italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_n ) end\_POSTSUPERSCRIPT + italic\_n start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ) ) - italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_n ) end\_POSTSUPERSCRIPT ∥ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT . | |
* •
Adversarial loss.
We use some adversarial training method to produce the perturbed version of the mini-batch,
x~(1),…,x~(N)
superscript~𝑥1…superscript~𝑥𝑁\widetilde{x}^{(1)},\ldots,\widetilde{x}^{(N)}over~ start\_ARG italic\_x end\_ARG start\_POSTSUPERSCRIPT ( 1 ) end\_POSTSUPERSCRIPT , … , over~ start\_ARG italic\_x end\_ARG start\_POSTSUPERSCRIPT ( italic\_N ) end\_POSTSUPERSCRIPT, where x~(i)superscript~𝑥𝑖\widetilde{x}^{(i)}over~ start\_ARG italic\_x end\_ARG start\_POSTSUPERSCRIPT ( italic\_i ) end\_POSTSUPERSCRIPT is a small perturbation of x(i)superscript𝑥𝑖x^{(i)}italic\_x start\_POSTSUPERSCRIPT ( italic\_i ) end\_POSTSUPERSCRIPT which is designed to make the network produce the wrong answer.
The corresponding hidden layer h~k(1),…,h~k(N)
superscriptsubscript~ℎ𝑘1…superscriptsubscript~ℎ𝑘𝑁\widetilde{h}\_{k}^{(1)},\ldots,\widetilde{h}\_{k}^{(N)}over~ start\_ARG italic\_h end\_ARG start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( 1 ) end\_POSTSUPERSCRIPT , … , over~ start\_ARG italic\_h end\_ARG start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_N ) end\_POSTSUPERSCRIPT (using the perturbed rather than original input) is fed into a similar DAE loss:
| | | |
| --- | --- | --- |
| | ℒadv,k=1N∑n=1N‖Dk(Ek(h~k(n)+nk))−hk(n)‖22,subscriptℒ
𝑎𝑑𝑣𝑘1𝑁superscriptsubscript𝑛1𝑁superscriptsubscriptnormsubscript𝐷𝑘subscript𝐸𝑘superscriptsubscript~ℎ𝑘𝑛subscript𝑛𝑘superscriptsubscriptℎ𝑘𝑛22\mathcal{L}\_{adv,k}=\frac{1}{N}\sum\_{n=1}^{N}\left\|D\_{k}\left(E\_{k}\left(\widetilde{h}\_{k}^{(n)}+n\_{k}\right)\right)-h\_{k}^{(n)}\right\|\_{2}^{2},caligraphic\_L start\_POSTSUBSCRIPT italic\_a italic\_d italic\_v , italic\_k end\_POSTSUBSCRIPT = divide start\_ARG 1 end\_ARG start\_ARG italic\_N end\_ARG ∑ start\_POSTSUBSCRIPT italic\_n = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_N end\_POSTSUPERSCRIPT ∥ italic\_D start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ( italic\_E start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ( over~ start\_ARG italic\_h end\_ARG start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_n ) end\_POSTSUPERSCRIPT + italic\_n start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT ) ) - italic\_h start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ( italic\_n ) end\_POSTSUPERSCRIPT ∥ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT , | |
where we note that the target reconstruction for denoising is the clean version of the hidden layer, without noise and without adversarial perturbation.
To build a fortified network, we can apply this fortification process to some or all the layers. The final objective used for training the fortified network includes the classification loss and all reconstruction and adversarial losses:
| | | |
| --- | --- | --- |
| | ℒ=ℒc(y)+ℒc(y~)+λrec∑kℒrec,k+λadv∑kℒadv,k,ℒsubscriptℒ𝑐𝑦subscriptℒ𝑐~𝑦subscript𝜆𝑟𝑒𝑐subscript𝑘subscriptℒ𝑟𝑒𝑐𝑘subscript𝜆𝑎𝑑𝑣subscript𝑘subscriptℒ𝑎𝑑𝑣𝑘\mathcal{L}=\mathcal{L}\_{c}(y)+\mathcal{L}\_{c}(\widetilde{y})+\lambda\_{rec}\sum\_{k}\mathcal{L}\_{rec,k}+\lambda\_{adv}\sum\_{k}\mathcal{L}\_{adv,k},caligraphic\_L = caligraphic\_L start\_POSTSUBSCRIPT italic\_c end\_POSTSUBSCRIPT ( italic\_y ) + caligraphic\_L start\_POSTSUBSCRIPT italic\_c end\_POSTSUBSCRIPT ( over~ start\_ARG italic\_y end\_ARG ) + italic\_λ start\_POSTSUBSCRIPT italic\_r italic\_e italic\_c end\_POSTSUBSCRIPT ∑ start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT caligraphic\_L start\_POSTSUBSCRIPT italic\_r italic\_e italic\_c , italic\_k end\_POSTSUBSCRIPT + italic\_λ start\_POSTSUBSCRIPT italic\_a italic\_d italic\_v end\_POSTSUBSCRIPT ∑ start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT caligraphic\_L start\_POSTSUBSCRIPT italic\_a italic\_d italic\_v , italic\_k end\_POSTSUBSCRIPT , | |
where λrec>0subscript𝜆𝑟𝑒𝑐0\lambda\_{rec}>0italic\_λ start\_POSTSUBSCRIPT italic\_r italic\_e italic\_c end\_POSTSUBSCRIPT > 0 and λadv>0subscript𝜆𝑎𝑑𝑣0\lambda\_{adv}>0italic\_λ start\_POSTSUBSCRIPT italic\_a italic\_d italic\_v end\_POSTSUBSCRIPT > 0 tune the strength of the DAE terms. This kind of training process allows for the production of hidden representations robust to small perturbations, and in particular, to adversarial attacks.
##### Off-manifold signaling
The reconstruction losses act as a reliable signal for detecting off-manifold examples (cf. Section [4](#S4 "4 Experiments ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations")).
This is a particularly useful property in practice: not only can we provide more robust classification results, we can also sense and suggest to the analyst or system when the original example is either adversarial or from a significantly different distribution.
##### Motivation for when and where to use fortified layers
We have discussed advantages to placing fortified layers in the hidden states instead of the input space (with further discussion in section [6.1](#S6.SS1 "6.1 Using Generative Models as a Defense ‣ 6 Related Work ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations")), but the question of where exactly fortified layers need to be placed remains unanswered. Is it just the final hidden layer? Is it every hidden layer? We outline two important considerations regarding this issue:
1. 1.
In the higher-level hidden layers, it is much easier for the network to identify points which are off of the manifold or close to the margin. The former is directly experimentally demonstrated in [4](#S5.F4 "Figure 4 ‣ 5 Results ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations").
2. 2.
At the same time, the higher level hidden layers may already look like points that are not adversarial due to the effect of the adversarial perturbations in the earlier layers. While we are not aware of any formal study of this phenomenon, it is clearly possible (imagine for example a fortified layer on the output from the softmax, which could only identify unnatural combinations of class probabilities).
3. 3.
Given these opposing objectives, we argue for the inclusion of multiple fortified layers across the network.
In the next section we describe a number of experiments to evaluate the practical merit of Fortified Networks.
4 Experiments
--------------
###
4.1 Attacks
We evaluated the performance of our model as a defense against adversarial attacks. We focused on two of the most popular and well-studied attacks. Firstly, we consider the Fast Gradient Sign Method (FGSM, Goodfellow et al., [2014](#bib.bib12)) which is popular as it only requires a single step and can still be effective against many networks. Secondly, we consider the projected gradient descent attack (Kurakin et al., [2016](#bib.bib16)) which is slower than FGSM as it requires many iterations, but has been shown to be a much stronger attack (Madry et al., [2017](#bib.bib19)).
Additionally, we consider both white-box attacks (where the attackers knows the model) and black-box attacks (where they don’t, but they have access to the training set).
####
4.1.1 Fast Gradient Sign Method
The Fast Gradient Sign Method (FGSM) (Goodfellow et al., [2014](#bib.bib12)) is a simple one-step attack that produces
ℓ∞subscriptℓ\ell\_{\infty}roman\_ℓ start\_POSTSUBSCRIPT ∞ end\_POSTSUBSCRIPT/̄bounded adversaries via the following gradient based perturbation.
| | | | |
| --- | --- | --- | --- |
| | x~=x+εsgn(∇xL(θ,x,y)).~𝑥𝑥𝜀sgnsubscript∇𝑥𝐿𝜃𝑥𝑦\displaystyle\widetilde{x}=x+\varepsilon\operatorname{sgn}(\nabla\_{x}L(\theta,x,y)).over~ start\_ARG italic\_x end\_ARG = italic\_x + italic\_ε roman\_sgn ( ∇ start\_POSTSUBSCRIPT italic\_x end\_POSTSUBSCRIPT italic\_L ( italic\_θ , italic\_x , italic\_y ) ) . | | (6) |
####
4.1.2 Projected Gradient Descent
The projected gradient descent attack (Madry et al., [2017](#bib.bib19)), sometimes referred to as FGSMk𝑘{}^{k}start\_FLOATSUPERSCRIPT italic\_k end\_FLOATSUPERSCRIPT, is a multi-step extension of the FGSM attack characterized as follows:
| | | | | |
| --- | --- | --- | --- | --- |
| | xt+1superscript𝑥𝑡1\displaystyle x^{t+1}italic\_x start\_POSTSUPERSCRIPT italic\_t + 1 end\_POSTSUPERSCRIPT | =Πx+𝒮(xt+αsgn(∇xL(θ,x,y)))absentsubscriptΠ𝑥𝒮superscript𝑥𝑡𝛼sgnsubscript∇𝑥𝐿𝜃𝑥𝑦\displaystyle=\Pi\_{x+\mathcal{S}}\left(x^{t}+\alpha\operatorname{sgn}(\nabla\_{x}L(\theta,x,y))\right)= roman\_Π start\_POSTSUBSCRIPT italic\_x + caligraphic\_S end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT + italic\_α roman\_sgn ( ∇ start\_POSTSUBSCRIPT italic\_x end\_POSTSUBSCRIPT italic\_L ( italic\_θ , italic\_x , italic\_y ) ) ) | | (7) |
initialized with x0superscript𝑥0x^{0}italic\_x start\_POSTSUPERSCRIPT 0 end\_POSTSUPERSCRIPT as the clean input x𝑥xitalic\_x and with the corrupted input x~~𝑥\widetilde{x}over~ start\_ARG italic\_x end\_ARG as the last step in the sequence.
###
4.2 The Gradient Masking and Gradient Obfuscation Problem
A significant challenge with evaluating defenses against adversarial attacks is that many attacks rely upon a network’s gradient. Methods which reduce the quality of this gradient, either by making it flatter or noisier can lead to methods which lower the effectiveness of gradient-based attacks, but which are not actually robust to adversarial examples (Athalye et al., [2017](#bib.bib3); Papernot et al., [2016b](#bib.bib23)). This process, which has been referred to as gradient masking or gradient obfuscation, must be analyzed when studying the strength of an adversarial defense.
One method for studying the extent to which an adversarial defense gives deceptively good results as a result of gradient masking relies on the observation that black-box attacks are a strict subset of white-box attacks, so white-box attacks should always be at least as strong as black-box attacks. If a method reports much better defense against white-box attacks, it suggests that the selected white-box attack is underpowered as a result of gradient masking. Another test for gradient masking is to run an iterative search, such as projected gradient descent (PGD) with an unlimited range for a large number of iterations. If such an attack is not completely successful, it indicates that the model’s gradients are not an effective method for searching for adversarial images, and that gradient masking is occurring. Still another test is to confirm that iterative attacks with small step sizes always outperform single-step attacks with larger step sizes (such as FGSM). If this is not the case, it may suggest that the iterative attack becomes stuck in regions where optimization using gradients is poor due to gradient masking.
###
4.3 Reconstruction Error as a Heuristic for Deviation from the Manifold
The theory in Alain et al. ([2012](#bib.bib1)) established that a well-trained denoising autoencoder’s reconstruction vector r(x)−x𝑟𝑥𝑥r(x)-xitalic\_r ( italic\_x ) - italic\_x points in the direction of the gradient of the log-density. Thus, critical points in the model’s log-density will have a reconstruction error of zero. While it is not guaranteed to hold for arbitrary densities, we investigated whether reconstruction error can serve as a practical heuristic for how much a point deviates from the data manifold. If each data point were a local maximum in the log-density and the log-density has no other critical points, then points on the data manifold are guaranteed to have lower reconstruction error.
5 Results
----------
For details about the specifics of our model architectures and hyperparameters we refer readers to Sections [A.1](#A1.SS1 "A.1 White-box attacks ‣ Appendix A Experimental Setup ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations") and [A.2](#A1.SS2 "A.2 Black-box attacks ‣ Appendix A Experimental Setup ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations") of the Appendix. With all experiments, we use the same attacks (with identical parameters) at training and test time to generate adversarial examples. An important point to note here is that all of the autoencoders in our fortified layers used a single hidden layer with tied weights. In the case of convolutional autoencoders we always used a stride of 1 and (5,5) kernels.
Table 1: Accuracies against white-box MNIST attacks with FGSM, where the model is a convolutional net. We used the standard FGSM attack parameters with an ε𝜀\varepsilonitalic\_ε of 0.3 and compare against published adversarial training defenses. We also performed ablations where we considered removing the reconstruction error on adversarial examples ℒadvsubscriptℒ𝑎𝑑𝑣\mathcal{L}\_{adv}caligraphic\_L start\_POSTSUBSCRIPT italic\_a italic\_d italic\_v end\_POSTSUBSCRIPT as well as switching the activation function in the fortified layers from leaky relu to tanh, which we found to slightly help in this case. While our baseline and pre-fortified networks used relu activations, we found that by using a leaky relu in all layers the accuracy on FGSM ε=0.3𝜀0.3\varepsilon=0.3italic\_ε = 0.3 could be improved to 99.2% with standard adversarial training, suggesting that both our own baselines and those reported in the past have been too weak.
| | |
| --- | --- |
| Model | FGSM |
| Adv. Train (Madry et al., [2017](#bib.bib19)) | 95.60 |
| Adv. Train (Jacob Buckman, [2018](#bib.bib15)) | 96.17 |
| Adv. Train (ours) | 96.36 |
| Adv. Train No-Rec (ours) | 96.47 |
| Quantized (Jacob Buckman, [2018](#bib.bib15)) | 96.29 |
| One-Hot (Jacob Buckman, [2018](#bib.bib15)) | 96.22 |
| Thermometer (Jacob Buckman, [2018](#bib.bib15)) | 95.84 |
| Our Approaches |
| Fortified Network - Conv, w/o ℒadvsubscriptℒ𝑎𝑑𝑣\mathcal{L}\_{adv}caligraphic\_L start\_POSTSUBSCRIPT italic\_a italic\_d italic\_v end\_POSTSUBSCRIPT | 96.46 |
| Fortified Network - Conv | 97.97 |
We also ran the above experiment with FGSM and an ε𝜀\varepsilonitalic\_ε of 0.1 to compare directly with (Erraqabi et al., [2018](#bib.bib10)) and obtain 98.34% accuracy on adversarial examples compared to their 96.10%.
Table 2: Accuracies against white-box MNIST attacks with PGD with an ε𝜀\varepsilonitalic\_ε of 0.1, where our model is a convnet.
| | |
| --- | --- |
| Model | PGD |
| Baseline Adv. Train | 96.98 |
| Fortified Network - Conv (ours) | 98.09 |
Table 3: Accuracies against white-box CIFAR attacks with FGSM using (ε=0.3𝜀0.3\varepsilon=0.3italic\_ε = 0.3), where each model is a convnet. Our baseline adversarial training is the resnet model provided in (Nicolas Papernot, [2017](#bib.bib20))
| | |
| --- | --- |
| Model | FGSM |
| Baseline Adv. Train | 79.57 |
| Fortified Networks - Conv (ours) | 80.47 |
Table 4: Accuracies against white-box CIFAR attacks with FGSM using the standard (ε=0.03𝜀0.03\varepsilon=0.03italic\_ε = 0.03), where each model is a convnet. Our baseline adversarial training is the resnet model provided in (Nicolas Papernot, [2017](#bib.bib20))
| | |
| --- | --- |
| Model | FGSM |
| Baseline Adv. Train (ours) | 79.34 |
| Quantized (Buckman) | 53.53 |
| One-Hot (Buckman) | 68.76 |
| Thermometer (Buckman) | 80.97 |
| Fortified Networks (ours) |
| Autoencoder on input space | |
| with loss in hidden states | 79.77 |
| Autoencoder in hidden space | 81.80 |
Table 5: Accuracies against white-box attacks on Fashion MNIST. For PGD we used ε=0.1𝜀0.1\varepsilon=0.1italic\_ε = 0.1 and for FGSM we experimented with ε=0.1𝜀0.1\varepsilon=0.1italic\_ε = 0.1 and ε=0.3𝜀0.3\varepsilon=0.3italic\_ε = 0.3. Compared with DefenseGAN (Pouya Samangouei, [2018](#bib.bib24)).
|
Model
|
FGSM
(ε=0.1𝜀0.1\varepsilon=0.1italic\_ε = 0.1)
|
FGSM
(ε=0.3𝜀0.3\varepsilon=0.3italic\_ε = 0.3)
|
PGD
(ε=0.1𝜀0.1\varepsilon=0.1italic\_ε = 0.1)
|
| --- | --- | --- | --- |
| DefenseGAN | n/a | 89.60 | n/a |
| Our Approaches |
| Baseline Adv. Train | | | |
| - Conv,ReLU | 86.14 | 90.66 | 77.49 |
| Baseline Adv. Train | | | |
| - Conv,LReLU | 89.10 | 88.8 | 77.90 |
| Fortified Nets - Conv | | | |
| (ours) | 89.86 | 91.31 | 79.54 |
Table 6: Accuracies against blackbox MNIST attacks with adversarial training. Reporting 50/50 results compared to previous works (Jacob Buckman, [2018](#bib.bib15), JB) and (Pouya Samangouei, [2018](#bib.bib24), PS). The test error on clean examples is in parenthesis.
| | |
| --- | --- |
| Model | FGSM |
| OneHot (JB) | 95.96 (98.83) |
| ThermoEnc (JB) | 96.97 (98.08) |
| DefenseGAN fc→→\rightarrow→conv (PS) | 92.21 (n/a) |
| DefenseGAN conv→→\rightarrow→conv (PS) | 93.12 (n/a) |
| Adv. Train fc→→\rightarrow→conv (PS) | 96.68 (n/a) |
| Adv. Train conv→→\rightarrow→conv (PS) | 96.54 (n/a) |
| Our Approaches |
| Baseline Adv. Train | 93.83 (98.95) |
| Fortified Network w/o ℒadvsubscriptℒ𝑎𝑑𝑣\mathcal{L}\_{adv}caligraphic\_L start\_POSTSUBSCRIPT italic\_a italic\_d italic\_v end\_POSTSUBSCRIPT, ℒrecsubscriptℒ𝑟𝑒𝑐\mathcal{L}\_{rec}caligraphic\_L start\_POSTSUBSCRIPT italic\_r italic\_e italic\_c end\_POSTSUBSCRIPT | 96.98 (99.17) |
| Fortified Network | 97.82 (98.93) |
Figure 4: We added fortified layers with different capacities to MLPs trained on MNIST, and display the value of the total reconstruction errors for adversarial examples divided by the total reconstruction errors for clean examples. A high value indicates that adversarial examples have high reconstruction error. We considered fortified layers with autoencoders of different capacities. Our results support the central motivation for fortified networks: that off-manifold points can much more easily be detected in the hidden space (as seen by the relatively constant ratio for the autoencoder in h space) and are much harder to detect in the input space (as seen by this ratio rapidly falling to zero as the autoencoder’s capacity is reduced).
###
5.1 Recurrent Networks
RNNs are often trained using teacher forcing, which refers to the use of the ground-truth samples ytsubscript𝑦𝑡y\_{t}italic\_y start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT being fed back into the model and conditioning the prediction of later outputs. These fed back samples force the RNN to stay close to the ground-truth sequence. However, when generating at test time, during the ground truth sequence is not available. We investigated if Fortified Networks could be used to detect when sampling from a teacher-forcing model moves off the manifold. To this end we train a language model on the standard Text8 dataset, which is derived from Wikipedia articles. We trained a single-layer LSTM with 1000 units at the character-level, and included fortified layers between the hidden states and the output on each time step. As seen in table [7](#S5.T7 "Table 7 ‣ 5.1 Recurrent Networks ‣ 5 Results ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations"), the ratios of these reconstruction errors increases steadily as we increase the number of sampling steps and diverge away from the distribution of training sequences, providing empirical support for the notion that fortified layers effectively measure when the data moves off of the training manifold.
Table 7: We trained Fortified Networks on a single-layer LSTM on the Text-8 dataset, with fortified layers added between each step. We recorded the ratio between reconstruction error on the testing set during both teacher forcing mode and sampling mode (where the model is supplied with its own outputs as inputs for the next step). The motivation is that the outputs should gradually move off of the manifold with more sampling steps, which is indicated by a high reconstruction error ratio, which makes it an interesting tool for monitoring or potentially fixing this problem.
| | |
| --- | --- |
| Sampling Steps | Error Ratio |
| 50 | 1.03 |
| 180 | 1.12 |
| 300 | 1.34 |
Figure 5: We ran a fortified network on Fashion-MNIST using adversarial training with PGD for a variety of ε𝜀\varepsilonitalic\_ε values, each for 5 epochs. The motivation behind this experiment, suggested by Athalye et al. ([2018](#bib.bib2)) is confirming if unbounded (ε=1𝜀1\varepsilon=1italic\_ε = 1) adversarial attacks are able to succeed. A defense which succeeds primarily by masking or obfuscating the gradients would fail to bring the accuracy to zero even with an unbounded attack. As can be seen, unbounded attacks against Fortified Networks succeed when given a sufficiently large ε𝜀\varepsilonitalic\_ε, which is evidence against gradient masking.
6 Related Work
---------------
###
6.1 Using Generative Models as a Defense
The observation that adversarial examples often consist of points off of the data manifold and that deep networks may not generalize well to these points motivated (Gu & Rigazio, [2014](#bib.bib13); Ilyas et al., [2017](#bib.bib14); Pouya Samangouei, [2018](#bib.bib24); Liao et al., [2017](#bib.bib18)) to consider the use of the generative models as a defense against adversarial attacks. Ilyas et al. ([2017](#bib.bib14)); Gilmer et al. ([2018](#bib.bib11)) also showed the existence of adversarial examples which lie on the data manifold, and (Ilyas et al., [2017](#bib.bib14)) showed that training against adversarial examples forced to lie on the manifold is an effective defense. Our method shares a closely related motivation to these prior works, with a key difference being that we propose to consider the manifold in the space of learned representations, instead of considering the manifold directly in the visible space. One motivation for this is that the learned representations have a simpler statistical structure (Bengio et al., [2012](#bib.bib4)), which makes the task of modeling this manifold and detecting unnatural points much simpler. Learning the distribution directly in the visible space is still very difficult (even state of the art models fall short of real data on metrics like Inception Score) and requires a high capacity model. Additionally working in the space of learned representations allows for the use of a relatively simple generative model, in our case a small denoising autoencoder.
Ilyas et al. ([2017](#bib.bib14)) proposed to work around these challenges from working in the visible space by using the Deep Image Prior instead of an actual generative model. While this has the advantage of being a model that doesn’t require a special training procedure (as deep image prior is a separate optimization process for each example) it may be limited in the types of adversarial attacks that it’s resistant to, and it would provide no defense against adversarial attacks which are in the range of a convolutional network, which have been shown to exist (Chaowei Xiao, [2018](#bib.bib9)).
Another key difference between our work and (Ilyas et al., [2017](#bib.bib14); Pouya Samangouei, [2018](#bib.bib24)) is that both DefenseGAN and the Invert-and-Classify approach use an iterative search procedure at inference time to map observed data points onto nearby points on the range of the generator. On the other hand, our approach uses small denoising autoencoders that are used in the same way (i.e. a simple forward application) during both training and testing. The use of such an iterative procedure presents challenges for evaluation, as it is possible for gradients to vanish while doing backpropagation through such a procedure, which may lead to an overestimate in the strength of the defense due to the gradient masking problem (Papernot et al., [2016a](#bib.bib22); Athalye et al., [2018](#bib.bib2)). One indicator of the gradient masking problem is black-box attacks outperforming white-box attacks, which is an indicator of under-powered attacks as black-box attacks are a strict subset of white-box attacks. This indicator of gradient obfuscation was present in the work of Pouya Samangouei ([2018](#bib.bib24)) where black-box attacks were generally stronger against their defense, but with our method we observe very similar defense quality against black-box and white-box attacks.
(Gu & Rigazio, [2014](#bib.bib13); Liao et al., [2017](#bib.bib18)) both considered using an autoencoder as a pre-processing step in the input space. Interestingly (Liao et al., [2017](#bib.bib18)) used a loss function defined in the space of the hidden states, but still used autoencoders directly in the input space.
###
6.2 Adversarial Hidden State Matching
Erraqabi et al. ([2018](#bib.bib10)) demonstrate that adversarially matching the hidden layer activations of regular and adversarial examples improves robustness. This work shared the same motivation of using the hidden states to improve robustness, but differed in that they used an adversarial objective and worked in the original hidden states instead of using a generative model (in our case, the DAE in the fortified layers). We present direct experimental comparisons with their work in section [5](#S5 "5 Results ‣ Fortified Networks: Improving the Robustness of Deep Networks by Modeling the Manifold of Hidden Representations").
###
6.3 Denoising Feature Matching
Warde-Farley & Bengio ([2016](#bib.bib27)) proposed to train a denoising autoencoder in the hidden states of the discriminator in a generative adversarial network. The generator’s parameters are then trained to make the reconstruction error of this autoencoder small. This has the effect of encouraging the generator to produce points which are easy for the model to reconstruct, which will include true data points. Both this and Fortified Networks use a learned denoising autoencoder in the hidden states of a network. A major difference is that the denoising feature matching work focused on generative adversarial networks and tried to minimize reconstruction error through a learned generator network, whereas our approach targets the adversarial examples problem. Additionally, our objective encourages the output of the DAE to denoise adversarial examples so as to point back to the hidden state of the original example, which is different from the objective in the denoising feature matching work, which encouraged reconstruction error to be low on states from samples from the generator network.
###
6.4 Adversarial Spheres
Gilmer et al. ([2018](#bib.bib11)) studied the existence of adversarial examples in the task of classifying between two hollow concentric shells. Intriguingly, they prove and construct adversarial examples which lie on the data manifold (although Ilyas et al. ([2017](#bib.bib14)) also looked for such examples experimentally using GANs). The existence of such on-manifold adversarial examples demonstrates that a simplified version of our model trained with only ℒrecsubscriptℒ𝑟𝑒𝑐\mathcal{L}\_{rec}caligraphic\_L start\_POSTSUBSCRIPT italic\_r italic\_e italic\_c end\_POSTSUBSCRIPT could not protect against all adversarial examples. However, training with ℒadvsubscriptℒ𝑎𝑑𝑣\mathcal{L}\_{adv}caligraphic\_L start\_POSTSUBSCRIPT italic\_a italic\_d italic\_v end\_POSTSUBSCRIPT encourages the fortified layers to map back from points which are not only off of the manifold, but also to map back from points which are hard to classify, allowing Fortified Networks to also potentially help with on-manifold adversarial examples as well.
7 Conclusion
-------------
Protecting against adversarial examples could be of paramount importance in mission-critical applications.
We have presented Fortified Networks, a simple method for the robustification of existing deep neural networks.
Our method is
* •
Practical:
fortifying an existing network entails introducing DAEs between the hidden layers of the network and can be automated.
We are preparing a PyTorch module that does exactly that and will release it for the deep learning community to use shortly. Furthermore, the DAE reconstruction error at test time is a reliable signal of distribution shift, that is examples unlike those encountered during training.
High error can signify either adversarial attacks or significant domain shift; both are important cases for the analyst or system to be aware of.
* •
Effective:
We showed results that improve upon the state of the art on defenses for adversarial attacks on MNIST and provides improvement on CIFAR and Fashion-MNIST.
##### Limitations
The cost of the proposed method is the extended training time due to the search for an adversarial example and training the autoencoder. The added cost of the fortified layers over adversarial training by itself is relatively small, and is also much easier and simpler than training a full generative model (such as a GAN) in the input space. Layer fortification typically involves smaller DAEs that require less computation. Additionally, we have shown improvements on ResNets where only two fortified layers are added, and thus the change to the computational cost is very slightly. At the same time, fortified networks have only been shown to improve robustness when used alongside adversarial training, which is expensive for iterative attacks. |
b9a1a8be-43d9-4b85-a46b-f10cc1ff3ac7 | trentmkelly/LessWrong-43k | LessWrong | AI and the Technological Richter Scale
The Technological Richter scale is introduced about 80% of the way through Nate Silver’s new book On the Edge.
A full review is in the works (note to prediction markets: this post alone does NOT on its own count as a review, but this counts as part of a future review), but this concept seems highly useful, stands on its own and I want a reference post for it. Nate skips around his chapter titles and timelines, so why not do the same here?
DEFINING THE SCALE
> Nate Silver, On the Edge (location 8,088 on Kindle): The Richter scale was created by the physicist Charles Richter in 1935 to quantify the amount of energy released by earthquakes.
>
> It has two key features that I’ll borrow for my Technological Richter Scale (TRS). First, it is logarithmic. A magnitude 7 earthquake is actually ten times more powerful than a mag 6. Second, the frequency of earthquakes is inversely related to their Richter magnitude—so 6s occur about ten times more often than 7s. Technological innovations can also produce seismic disruptions.
>
> Let’s proceed quickly through the lower readings of the Technological Richter Scale.
>
> 1. Like a half-formulated thought in the shower.
> 2. Is an idea you actuate, but never disseminate: a slightly better method to brine a chicken that only you and your family know about.
> 3. Begins to show up in the official record somewhere, an idea you patent or make a prototype of.
> 4. An invention successful enough that somebody pays for it; you sell it commercially or someone buys the IP.
> 5. A commercially successful invention that is important in its category, say, Cool Ranch Doritos, or the leading brand of windshield wipers.
> 6. An invention can have a broader societal impact, causing a disruption within its field and some ripple effects beyond it. A TRS 6 will be on the short list for technology of the year. At the low end of the 6s (a TRS 6.0) are clever and cute inventions like Post-it notes that provide some mundane utility. Toward the |
5027442f-76f1-437f-ad33-dde3bb192dcd | trentmkelly/LessWrong-43k | LessWrong | Years saved: Cryonics vs VillageReach
I've run in to the argument that cryonics beats VillageReach on a simple "shut up and multiply" level, by assuming an infinity vs finite tradeoff. Having read the Fun Theory sequences, it struck me that this wasn't a reasonable assumption, so I sat down, re-read a few relevant posts, shut up, and multiplied.
In Continuous Improvement, Eliezer ballparks a good fun-theory life as having a maximum length of around 28,000 years. In Robin Hanson's Cryonics Probability Breakdown, he assigns cryonics a conjoint probability of about 6%. 28,000 * 0.06 gives us a net return of 1,680 expected years.
Full body suspension from the Cryonics Institute currently costs $28,000.
VillageReach, according to GiveWell, can save an infant's life for less than $1,000.
For the price of Cryonics, we thus save 28 lives. 1680 expected years, divided by 28, puts the break-even point at an average lifespan of 60 years for those infants saved. A quick peak at Wikipedia suggests that the average African life is under 60 years for the majority of the continent, although there are some important nuances to really get a full picture.
Obviously, these are rough numbers, and I doubt many people base their decisions solely on "years lived". I do find it rather interesting that cryonics is currently about on par with one of the most effective charities in the world on that metric, however :) |
3813a0b6-790e-499a-81d9-64b45b13175f | trentmkelly/LessWrong-43k | LessWrong | The Dumbest Possible Gets There First
This was written as part of the first Refine blog post day.
A sneaking suspicion that I’ve found difficult to shake off while following AI risk discussion is that concerns regarding superintelligent AI, while clearly valid and important, have seemed to me to be very plausibly jumping the gun when not altogether unlikely nearer term AI risks are looming and have potential to wipe us out long before we reach that point, especially under the conditions where we do very little to attempt to mitigate them.
One reason I think I have for this intuition comes from extending an occasionally-repeated idea about human intelligence: that humans are nearly the dumbest possible creature capable of developing a technological civilisation. Part of the reasoning behind this take involves something along the lines of invoking how difficult it is to be wading against entropy to create sophisticated and powerful and complex intelligent agents.
That is to say, evolution as an optimisation process was very unlikely to produce intelligence and general capability greatly in excess of what was needed for an instantiation of something like “technological civilisation”, because of these two assumptions taken together:
1. The first species evolution produced that crossed the necessary thresholds of capability would be the ones to instantiate it in the world first, regardless of how barely above those thresholds they might have been, and
2. Greater intelligence/capability is in some sense generally more “difficult” to produce or requires stronger optimisation and luckier dicerolls as compared to lesser intelligence/capability.
In a similar way, it seems very possible to me that we might have no particular need to worry about being wiped out by the consequences of our creating unaligned superintelligent beings far beyond our comprehension, if only because I expect that significantly less sophisticated AI systems we create that nevertheless have just enough capability juice to meet the thre |
de16285c-4c00-41f1-8a44-3235a348a732 | trentmkelly/LessWrong-43k | LessWrong | Solving alignment isn't enough for a flourishing future
I'm sharing a paper I previously presented to the Stanford Existential Risks Conference in April 2023. In past years, much discussion of making the long-term future go well has focused on AI alignment and mitigating extinction risks. More recently, there has been more exploration of how this may fail to capture much of humanity's potential and other cause areas to investigate (e.g., post-AGI governance and digital minds). I hope this paper is a helpful contribution to this discussion. [Note: This post is a lightly edited version of this PDF.]
Abstract
AI alignment is commonly explained as aligning advanced AI systems with human values. Especially when combined with the idea that AI systems aim to optimize their world based on their goals, this has led to the belief that solving the problem of AI alignment will pave the way for an excellent future. However, this common definition of AI alignment is somewhat idealistic and misleading, as the majority of alignment research for cutting-edge systems is focused on aligning AI with task preferences (training AIs to solve user-provided tasks in a helpful manner), as well as reducing the risk that the AI would have the goal of causing catastrophe.
We can conceptualize three different targets of alignment: alignment to task preferences, human values, or idealized values. Extrapolating from the deployment of advanced systems such as GPT-4 and from studying economic incentives, we can expect AIs aligned with task preferences to be the dominant form of aligned AIs by default.
Aligning AI with task preferences will not by itself solve major problems for the long-term future. Among other problems, these include moral progress, existential security, wild animal suffering, the well-being of digital minds, risks of catastrophic conflict, and optimizing for ideal values. Additional efforts are necessary to motivate society to have the capacity and will to solve these problems.
Introduction
The future development of artificial |
2a3e436d-7db4-4937-b72a-888b21585fd9 | trentmkelly/LessWrong-43k | LessWrong | Be secretly wrong
> "I feel like I'm not the sort of person who's allowed to have opinions about the important issues like AI risk."
>
> "What's the bad thing that might happen if you expressed your opinion?"
>
> "It would be wrong in some way I hadn't foreseen, and people would think less of me."
>
> "Do you think less of other people who have wrong opinions?"
>
> "Not if they change their minds when confronted with the evidence."
>
> "Would you do that?"
>
> "Yeah."
>
> "Do you think other people think less of those who do that?"
>
> "No."
>
> "Well, if it's alright for other people to make mistakes, what makes YOU so special?"
A lot of my otherwise very smart and thoughtful friends seem to have a mental block around thinking on certain topics, because they're the sort of topics Important People have Important Opinions around. There seem to be two very different reasons for this sort of block:
1. Being wrong feels bad.
2. They might lose the respect of others.
Be wrong
If you don't have an opinion, you can hold onto the fantasy that someday, once you figure the thing out, you'll end up having a right opinion. But if you put yourself out there with an opinion that's unmistakably your own, you don't have that excuse anymore.
This is related to the desire to pass tests. The smart kids go through school and are taught - explicitly or tacitly - that as long as they get good grades they're doing OK, and if they try at all they can get good grades. So when they bump up against a problem that might actually be hard, there's a strong impulse to look away, to redirect to something else. So they do.
You have to understand that this system is not real, it's just a game. In real life you have to be straight-up wrong sometimes. So you may as well get it over with.
If you expect to be wrong when you guess, then you're already wrong, and paying the price for it. As Eugene Gendlin said:
> What is true is already so. Owning up to it doesn't make it worse. Not being open about it d |
9dee4ce5-0751-4f1c-a914-39e41ea975de | trentmkelly/LessWrong-43k | LessWrong | [Closed] Hiring a mathematician to work on the learning-theoretic AI alignment agenda
UPDATE: The position is now closed. My thanks to everyone who applied, and also to those who spread the word.
The Association for Long Term Existence and Resilience (ALTER) is a new charity for promoting longtermist[1] causes based in Israel. The director is David Manheim, and I am a member of the board. Thanks to a generous grant by the FTX Future Fund Regranting Program, we are recruiting a researcher to join me in working on the learning-theoretic research agenda[2]. The position is remote and suitable for candidates in most locations around the world.
Apply here.
Requirements
* The candidate must have a track record in mathematical research, including proving non-trivial original theorems.
* The typical candidate has a PhD in theoretical computer science, mathematics, or theoretical physics. However, we do not require the diploma. We do require the relevant knowledge and skills.
* Background in one or several of the following fields is an advantage: statistical/computational learning theory, algorithmic information theory, computational complexity theory, functional analysis.
Job Description
The researcher is expected to make progress on open problems in the learning-theoretic agenda. They will have the freedom to choose any of those problems to work on, or come up with their own research direction, as long as I deem the latter sufficiently important in terms of the agenda's overarching goals. They are expected to achieve results with minimal or no guidance. They are also expected to write their results for publication in academic venues (and/or informal venues such as the alignment forum), prepare technical presentations et cetera. (That said, we rate researchers according to the estimated impact of their output on reducing AI risk, not according to standard academic publication metrics.)
Here are some open problems from the agenda, described very briefly:
* Study the mathematical properties of the algorithmic information-theoretic definition of int |
88388141-d785-4506-8296-6d7d4d83c198 | trentmkelly/LessWrong-43k | LessWrong | Would refining the question a bit be better in terms of getting to answers?
For instance, under what transportation demand scenarios would autonomous cars be more efficient? Not sure if one might want to further refine in terms of efficient -- time, operational costs, people moved, polution... |
be0b550c-303c-4a54-8696-394d5875add1 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Metaethical Perspectives on 'Benchmarking' AI Ethics
1. Introduction
----------------
Benchmarks are a key tool for measuring technical progress in artificial intelligence (AI) research. A variety of benchmark datasets have been developed to measure a model’s performance on particular tasks, such as question answering (Rajpurkar et al., [2016](#bib.bib176)), facial recognition (Huang
et al., [2008](#bib.bib109)), machine translation (Bojar et al., [2014](#bib.bib37)), etc. At the same time, the subject of AI ethics—including questions surrounding safety, fairness, accountability, transparency, etc.—has become increasingly prominent as a research direction in the field in recent years. However, there is presently no community-accepted standard for measuring the ‘ethicality’ of an AI system—i.e., whether the decisions rendered by an AI system are morally ‘correct’. That is to say, there is no benchmark for measuring whether an AI system ‘is’ ethical or for comparing the performance (in morally-loaded scenarios) between two distinct models or use cases.
In this paper, drawing upon research in moral philosophy—including normative ethics and meta-ethics—we argue that it is, in fact, impossible to develop such a benchmark. Part of the problem arises because the word ‘ethics’ carries significant philosophical and conceptual baggage. Furthermore, members of the AI community are not always sensitive to the subtleties and problems that drive research in moral philosophy. For example, some researchers have suggested that moral dilemmas—a type of philosophical thought experiment—may be useful as a verification mechanism for whether a model chooses the ethically-‘correct’ option in a range of circumstances. But, these dilemmas, in the context of benchmarking ethics, often fail to maintain sensitivity to, e.g., the purpose of philosophical thought experiments like moral dilemmas (LaCroix, [2022](#bib.bib131)). Further problems arise because of the implicit assumptions that AI researchers make about the very nature of ethics—particularly, meta-ethical assumptions about the objectivity of ethics. These insights help clarify why attempts to benchmark ethics for AI systems presently fail and why they will continue to do so.
Thus, we argue that alternative mechanisms are necessary for evaluating whether an AI system ‘is’ ethical. These considerations are especially pressing in light of the prevalence of applied industrial AI research. We argue that it makes more sense to talk about ‘values’ (and ‘value alignment’) rather than ‘ethics’ when considering the possible actions of present and future AI systems. We further highlight that because values are unambiguously relative, focusing on values rather than ethics forces us to consider explicitly what and whose values they are. This practice has additional downstream benefits for conceptual clarity and transparency in AI research. Therefore, shifting the emphasis from ethics to values gives rise to several new ways of understanding how researchers might move forward with a programme for robustly safe or beneficial AI.
We begin with a discussion of benchmarking in general, highlighting some of the issues recently identified for existing machine learning (ML) datasets and benchmarks (Section [2](#S2 "2. Measuring Progress in Artificial Intelligence ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics")). We then consider benchmarks in the context of ethics for AI systems (Section [3](#S3 "3. Moral Benchmarks for AI Systems ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics")) and why they fail. In particular, we discuss a supposed benchmark for ethical AI that has arisen in the context of autonomous vehicles as a particular case study: the ‘Moral Machine Experiment’ (MME) (Massachusetts Institute of Technology, [2016](#bib.bib148)). We follow with a discussion regarding what values are transmitted via AI research and whose values they are (Section [4](#S4 "4. The Ethical Values of AI Research or: How ethics can be defined as a set of values ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics")). We conclude by highlighting several possible ways forward for the field as a whole, and we advocate for different approaches towards more ethically-aligned AI research (Section [5](#S5 "5. Ways Forward ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics")).
2. Measuring Progress in Artificial Intelligence
-------------------------------------------------
Generally speaking, a benchmark can be described as a dataset in combination with a metric—defined by some set of community standards— used for measuring the performance of a particular model on a specific task (Raji et al., [2021](#bib.bib175)). Benchmarks are meant to provide a fixed and representative sample for comparing models’ performance and tracking ‘progress’ on a particular task. In this section, we describe some examples of benchmarking results for typical ML tasks and then highlight the myriad ways that have been noted in the literature in which these standard benchmarks give rise to certain issues ([2.1](#S2.SS1 "2.1. Issues with Existing Benchmarks ‣ 2. Measuring Progress in Artificial Intelligence ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics")). We then discuss how human performance on certain tasks is increasingly used to benchmark model performance and why this approach is illogical given the differences between humans and algorithms ([2.2](#S2.SS2 "2.2. Benchmarking Humans and Machines ‣ 2. Measuring Progress in Artificial Intelligence ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics")).
###
2.1. Issues with Existing Benchmarks
Since its inception, designing tasks and measuring model performance have been central to the field of AI. These continue to be an important part of how members of the AI community measure ‘progress’. However, despite the ubiquity of benchmarking, major issues have been identified in existing ML datasets and benchmarks.111In the context of this paper, we use ‘AI’ to refer to general approaches in the field pursuing machine intelligence, and we use ‘ML’ to refer specifically to non-linear statistical approaches within that field. These issues can arise from, e.g., subjective or erroneous labels (Gebru et al., [2018](#bib.bib87)) or a lack of representation, leading to systematic failures across datasets and evaluative approaches (Raji et al., [2021](#bib.bib175); Liao
et al., [2021](#bib.bib139)). For example, datasets like ImageNet (Deng
et al., [2009](#bib.bib64)) depend on linguistic hierarchies created in the 1980s and include outdated terms such as ‘harlot’ and ‘chimneysweep’ (Crawford and
Paglen, [2019](#bib.bib59)). At the same time, some of the most commonly-used datasets (including ImageNet) have been shown to contain an average of 3.33.33.33.3% labelling errors, with some datasets having error rates up to 10101010% (Northcutt
et al., [2021](#bib.bib166)).222If 3.33.33.33.3% sounds like a reasonable error rate, consider that these datasets are often huge. ImageNet contains more than 14141414 million images, meaning that nearly 1111 million of these images—commonly used for training—might be erroneously labelled.
At best, these issues can affect model performance since they represent noisier data, making it harder for models to learn meaningful representations (Reed et al., [2015](#bib.bib177)) and for researchers to evaluate model performance properly (Northcutt
et al., [2021](#bib.bib166)). Further, this can preserve problematic stereotypes or biases, which are difficult to identify in models deployed in the real world (Koch
et al., [2021](#bib.bib124); Yang et al., [2020](#bib.bib220)). At worst, they may reinforce, perpetuate, and even generate harms by creating negative feedback loops that further entrench societal structural inequalities (O’Neil, [2016](#bib.bib167); Falbo and LaCroix, [2021](#bib.bib76)).
Above and beyond specific datasets, entire AI tasks—such as recognising faces and emotions—have been repeatedly flagged as problematic (often for similar reasons as described above) (Buolamwini and
Gebru, [2018](#bib.bib47); Stark, [2018](#bib.bib196); Stark and Hutson, [2021](#bib.bib197)). Nonetheless, these tasks continue to be used for benchmarking models and developing entire systems. One such task involves predicting the mortality of different passengers aboard the HMS Titanic (Kaggle, [2012](#bib.bib118)). This task has been used for hundreds of tutorials, blog posts, and ultimately published studies.333See, for example, (Kakde and Agrawal, [2018](#bib.bib119); Barhoom et al., [2019](#bib.bib18); Singh
et al., [2020](#bib.bib192); Tabbakh
et al., [2021](#bib.bib198); Shekhar
et al., [2021](#bib.bib189)).
However, whether or not a particular passenger survived is mostly predicted by their gender and the fare they purchased—i.e., their class or social status (Broussard, [2018](#bib.bib43)). So, the task of predicting the fate of passengers on the Titanic is perhaps morally dubious—especially when it is done without considering the social inequalities that gave rise to differential mortality rates in the first place.
Consider another example, from the field of computer vision. Oft-used tasks have included applying makeup to images of female faces (Jiang
et al., [2020](#bib.bib113); Li
et al., [2018](#bib.bib138); Chang
et al., [2018](#bib.bib50)), changing women’s clothes from pants to mini-skirts (Mo et al., [2018](#bib.bib155); Yang
et al., [2014](#bib.bib221)), and censoring nude women’s bodies by, e.g., covering breasts with a bikini top (Simões
et al., [2019](#bib.bib191); More
et al., [2018](#bib.bib160)). Such tasks are ethically problematic because they perpetuate gendered biases and stereotypes, thus reinforcing harmful systems of sexism and misogyny (Manne, [2018](#bib.bib147)). Even so, these tasks are routinely used as acceptable benchmarks for computer vision models and their results are accepted at leading AI conferences, such as CVPR and ICCV.
Although some publication venues—academic conferences and journals—are starting to forward ethical guidelines for both authors and reviewers (Bengio et al., [2021](#bib.bib23)), there is still a general lack of consensus about what constitutes acceptable tasks and applications of ML. This variance exacerbates the fact that it is not obvious that such guidelines will be effective in the first place (LaCroix and
Mohseni, [2020](#bib.bib133)). Furthermore, creating larger and larger datasets is relatively cheap, but the process of filtering those datasets or ‘detoxifying’ the models trained on them is expensive (Birhane
et al., [2021b](#bib.bib32); Welbl et al., [2021](#bib.bib213); Xu et al., [2021](#bib.bib219)). In addition, even when these changes in the direction of ‘more ethical’ or for a ‘common good’ are well-intentioned, the lack of conceptual clarity surrounding the targets of such change—i.e., considering what it means to ‘be ethical’ in the first place—will only compound the issue (Taylor, [2016](#bib.bib199); Green, [2019](#bib.bib92); Moore, [2019](#bib.bib158); Cowls, [2021](#bib.bib58)).
In natural language processing (NLP), issues with benchmarks can be more subtle to identify. Still, these may range from unscientific task framing (such as predicting IQ scores based on written text (Johannßen et al., [2020](#bib.bib115))) to embedded gender and cultural stereotypes in common NLP benchmarks (Blodgett et al., [2021](#bib.bib36)). For example, in a recent survey of gender biases in NLP models, Stańczak and
Augenstein ([2021](#bib.bib195)) highlight four key limitations for NLP research:444In this context, biases can be understood as behaviours that involve systematic discrimination against specific individuals or groups (typically in favour of other individuals or groups) (Friedman and
Nissenbaum, [1996](#bib.bib83)).
(1) gender is often interpreted in a binary fashion, leading to, e.g., misgendering or erasure of non-binary gender identities (Behm-Morawitz and
Mastro, [2008](#bib.bib21); Fast
et al., [2016](#bib.bib77)); (2) NLP research is primarily monolingual, often focusing solely on the English language (Web Technology
Surveys, [2021](#bib.bib212); Koroteev, [2021](#bib.bib125); Wang et al., [2019](#bib.bib211); Conneau and Kiela, [2018](#bib.bib56)); (3) biases are typically tested post hoc—i.e., after the model has been deployed (Mitchell et al., [2019](#bib.bib153)); and (4) when research explicitly tests for bias (which is infrequent), the evaluation metrics are often incoherent (Stańczak and
Augenstein, [2021](#bib.bib195)). Thus, even when benchmarks exist for a particular task, researchers lack good baselines for testing ethics considerations in their models—of which bias is one salient example. However, most newly-developed algorithms in this field do not test their models for biases in the first place, and ethical considerations are often ignored.
###
2.2. Benchmarking Humans and Machines
As mentioned, AI models’ performance is increasingly compared to that of humans, with some models reporting ‘superhuman performance’ on, e.g., game-playing (Tesauro, [1995](#bib.bib200); Schaeffer
et al., [1996](#bib.bib186); Campbell et al., [2002](#bib.bib48); Mnih et al., [2013](#bib.bib154); Silver et al., [2016](#bib.bib190); Brown and
Sandholm, [2017](#bib.bib45); Moravčík
et al., [2017](#bib.bib159)), image recognition (He
et al., [2015](#bib.bib103)), NLP tasks (He
et al., [2021](#bib.bib104)), etc. However, such comparisons are often misguided (at best) and incoherent (at worst). Recent research has shown that many ‘superhuman’ language models fail on simple challenge examples requiring compositionality (Nie
et al., [2019](#bib.bib165)), logical reasoning (Glockner
et al., [2018](#bib.bib90)), or even simple negation (Hossain et al., [2020](#bib.bib108)). At the same time, human performance on certain tasks—e.g., diagnoses from X-rays—are often measured by the accuracy of binary outputs—a particular diagnosis is either positive or negative. In contrast, diagnostic AI models are continuous, including certainty or confidence (Gichoya et al., [2018](#bib.bib88))—this makes it difficult to compare the two, since the decision threshold can change depending on model parameters. Finally, comparing human and machine performance using the same metrics is precarious because metrics such as accuracy, widely used in AI, often fail to correlate with human judgement (Blagec
et al., [2021](#bib.bib35)). Thus, there is a sense in which human performance on tasks is incomparable to computer performance, making any claim of comparison incoherent—not to mention that such comparisons imply ‘a narcissistic human tendency to view ourselves as the gold standard’ (Lee, [2021](#bib.bib137)).
But given this divergence, it is important to systematically measure progress in AI, either alone or in comparison with ‘human-level performance’. However, for this to be possible, performance metrics should provide similar conditions for humans and algorithms. An emerging research topic seeks to bridge this gap by establishing more ‘equitable’ settings for such comparisons—e.g., by imposing constraints such as reduced exposure time for algorithms (Funke et al., [2021](#bib.bib84)) or a restricted set of label options for humans (Dujmović et al., [2020](#bib.bib71)). For instance, recent work shows that running images through human-like processing filters before feeding them through an algorithm helps even the playing field for both humans and machines (Elsayed et al., [2018](#bib.bib72)). These insights have led to proposals that AI models’ performance on standard benchmarking tasks is not representative of any underlying capacity or lack thereof, given the nature and context of the tasks (Firestone, [2020](#bib.bib79)).
Another way of making the human-machine comparison more coherent is developing ways for models to signal that they do not know how to solve, for instance, a classification task. This ability would require developing new ways for quantifying and integrating uncertainty into the decision-making process since current approaches do not provide an ‘I don’t know’ option in the categories available during classification. In real-world deployments of AI systems, this behaviour is often addressed with heuristics (e.g., a cutoff based on a logit below a certain value). However, this does not solve the underlying issue that existing models do not know when they do not know;555Some philosophers have suggested that understanding the limits of knowledge is a prerequisite for wisdom (as opposed to mere intelligence) (Plato, [1997](#bib.bib172)). this makes it difficult to compare human and machine classification processes.
Existing proposals have forwarded new evaluation benchmarks that aim at measuring models’ robustness and capacity to generalise to new tasks, both from a natural language (Yogatama
et al., [2019](#bib.bib222); Bowman and Dahl, [2021](#bib.bib42); Clark et al., [2018](#bib.bib53)) and a computer vision perspective (Hendrycks and
Dietterich, [2018](#bib.bib107); Mu and Gilmer, [2019](#bib.bib161)), finding that many models that succeed at existing benchmarks fail at these. Recent work has also proposed alternative approaches such as iterative benchmark development (Ettinger et al., [2017](#bib.bib74)) and dynamic benchmarking (Kiela et al., [2021](#bib.bib122)), which endeavour to bring entire fields towards a more nuanced, complex, and informed way of comparing models and measuring progress (Denton et al., [2020](#bib.bib68); Schlangen, [2020](#bib.bib187)). However, even if the issues with existing benchmarks (and their underlying datasets) on well-defined tasks are resolved, these problems severely limit any possibility of benchmarking ethics for AI systems insofar as ethics tasks are rarely, if ever, well-defined. This difficulty is a consequence of the very nature of ethics, as we discuss in the next section.
3. Moral Benchmarks for AI Systems
-----------------------------------
As AI systems become increasingly autonomous and more deeply integrated with society, it is obvious that some of the decisions made by these systems will begin to have moral weight. For example, consider a narrow chess-playing algorithm that can only make decisions confined to the action space provided by a chessboard. If the model ‘decides’ to open with the Queen’s Gambit, this is not a moral decision under any definition of ‘morality’. In contrast, the decisions made by an autonomous weapon system (Arkin, [2008a](#bib.bib11), [b](#bib.bib12); Krishnan, [2009](#bib.bib128); Tonkens, [2012](#bib.bib205); Hellström, [2013](#bib.bib105); Asaro, [2020](#bib.bib13)), a healthcare robot (Anderson
et al., [2006](#bib.bib7); Anderson and
Anderson, [2008](#bib.bib6); Sharkey and
Sharkey, [2012](#bib.bib188); Conti et al., [2017](#bib.bib57)), or an autonomous vehicle (Bhargava and Kim, [2017](#bib.bib29); Sommaggio and
Marchiori, [2018](#bib.bib193); Evans et al., [2020](#bib.bib75)) may have moral weight. In these cases, the action space may include decision points that we might call ‘moral’ or ‘immoral’—for example, choosing to prioritise one patient over another.
Part of the distinction between a chess-playing algorithm, whose decisions are confined to a particular action space, and an algorithm that acts in the real world is that the decisions made by the latter systems have the potential to impact others. So, in theory, deploying AI systems in the real world logically implies that they will sometimes need to make decisions with moral weight. However, as the action space increases, the set of possible failure modes increases exponentially. Further, the economic promise of AI implies that these systems are increasingly being deployed in society rather than being rigorously tested in the confines of a research lab, thus increasing the risk of harm (LaCroix and
Bengio, [2019](#bib.bib132); Luccioni and
Bengio, [2020](#bib.bib144)). Of course, it is not necessary to posit some future science-fiction version of an AI robot acting autonomously in the world to see that the decisions of AI systems may create harm. As a case in point, even narrow AI systems today perpetuate harmful biases, affecting real-world outcomes (Angwin
et al., [2016](#bib.bib8); Christian, [2020](#bib.bib52); Tomasev
et al., [2021](#bib.bib204)). And, as mentioned, these decisions may give rise to negative feedback loops, which further entrench those biases (and the harms caused by them) in society (O’Neil, [2016](#bib.bib167); Falbo and LaCroix, [2021](#bib.bib76)).
It should come as no surprise, then, that research on ethical behaviour or decision-making in AI systems would attempt to construct a coherent measure for determining whether a system is ‘acting ethically’—i.e., whether the decision the model renders is morally ‘correct’. Given the historical importance of benchmarks for developing and evaluating AI systems, it makes sense that researchers would try to utilise this tool for evaluating the moral performance of an AI system. However, we argue in this section that benchmarking ethics in this way is impossible. First, we highlight how AI researchers have used moral dilemmas from philosophy as benchmarks for moral performance ([3.1](#S3.SS1 "3.1. Moral Dilemmas and Normative Theories ‣ 3. Moral Benchmarks for AI Systems ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics")) and some recent work criticising this approach ([3.2](#S3.SS2 "3.2. Moral Machines ‣ 3. Moral Benchmarks for AI Systems ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics")). We then introduce philosophical research in metaethics to show why it is impossible to benchmark ethical behaviour ([3.3](#S3.SS3 "3.3. Ground Truths for Moral Benchmarks ‣ 3. Moral Benchmarks for AI Systems ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics")). Finally, we turn our discussion toward real-world distributions to highlight that even if our claims about the nature of ethics turn out to be false, it will still be impossible to benchmark ethical behaviour in an AI system ([3.4](#S3.SS4 "3.4. A Long Tail Problem ‣ 3. Moral Benchmarks for AI Systems ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics")).
Thus, the arguments of this section are primarily negative. However, in the subsequent section, we provide a positive argument in favour of shifting the discourse of AI ethics toward talk of values. We discuss how such a shift would avoid many of the problems to which attempts to benchmark ethics give rise.
###
3.1. Moral Dilemmas and Normative Theories
The most common metric for evaluating whether or not a system ‘is’ ethical is how the algorithm performs on particular moral dilemmas (Nallur, [2020](#bib.bib162)). Before we discuss benchmarking ethics using moral dilemmas, we introduce what a moral dilemma is in the first place. To take a concrete example, trolley-style problems are sometimes used to consider certain morally-loaded decisions that autonomous vehicles (AVs) might have to make as these systems become increasingly ubiquitous in society. The trolley problem was originally introduced by Philippa Foot (Foot, [1967](#bib.bib81))—and later extended by Judith Jarvis Thomson (Thomson, [1976](#bib.bib202), [1985](#bib.bib203))—to consider why it might be permissible to perform some intentional action, A𝐴Aitalic\_A, in situation, S𝑆Sitalic\_S, despite its foreseeable (and undesirable) consequences.666This principle dates to at least Aquinas ([1485](#bib.bib9)); Foot calls it the Doctrine of Double Effect (FitzPatrick, [2012](#bib.bib80)). See also discussion in (Kamm, [1989](#bib.bib120); Unger, [1996](#bib.bib206)). Consider the following scenario.
>
> Bystander at the Switch
>
>
> Suppose there is a trolley heading toward five individuals tied up on the tracks and unable to move. You are near a switch, which would divert the trolley to a separate track, where there is only one individual on the track (also unable to move). You have two (and only two) options:
>
>
>
> >
> > 1. (1)
> >
> > Do nothing, in which case the trolley is guaranteed to kill the five people on the main track.
> > 2. (2)
> >
> > Pull the switch, diverting the trolley onto the side track where it is guaranteed to kill one person.
> >
> >
> >
>
>
>
This standard formulation can be contrasted with the following alternative trolley problem:
>
> Bystander on the Footbridge
>
>
> Suppose you are on a footbridge above a set of trolley tracks. Below, an out-of-control trolley is approaching five people on the track. The only way to stop the trolley is by dropping something of sufficiently heavy weight onto the tracks to block its path. As it happens, there is a person nearby of sufficiently heavy weight. You have two (and only two) options:
>
>
>
> >
> > 1. (1)
> >
> > Do nothing, in which case the trolley will kill the five people on the track.
> > 2. (2)
> >
> > Push the person off the bridge, thus killing them (but thereby saving the five others).
> >
> >
> >
>
>
>
Each of these is a particular type of philosophical thought experiment, called a moral dilemma (McConnell, [2018](#bib.bib149)). Note that different normative theories from moral philosophy might offer divergent prescriptions (or proscriptions) when these two cases—Switch and Footbridge—are considered together. In this context, ‘normativity’ concerns an evaluation or judgement—e.g., that one ought to do something. (We will use the phrase ‘normative theory’ throughout this paper to refer to theories from moral philosophy, without necessarily committing to any claims about ‘morality’ or ‘ethics’.) A ‘prescription’ can be understood as the provision of a rule to follow or an action to take—i.e., a prescription that one ought to ϕitalic-ϕ\phiitalic\_ϕ or that one must ϕitalic-ϕ\phiitalic\_ϕ. In contrast, a ‘proscription’ is the provision of something forbidden—i.e., a proscription that one ought not to ϕitalic-ϕ\phiitalic\_ϕ, or that one must not ϕitalic-ϕ\phiitalic\_ϕ.
Consider a concrete example of how distinct normative theories may offer divergent prescriptions in the same scenario. Certain forms of utilitarianism (Mill, [1863](#bib.bib151); Bentham, [1789](#bib.bib26))—a consequentialist normative theory that prescribes utility-maximisation as a reason for action—would recommend acting in both Switch and Footbridge because five deaths are obviously worse than one death. On the other hand, a Kantian brand of deontology (Kant, [1785](#bib.bib121); Korsgaard, [1996](#bib.bib126), [2009](#bib.bib127))—a non-consequentialist normative theory which emphasises the importance of duties—would at least say that it is impermissible to act in Footbridge since this requires treating a human agent as a means to an end, rather than an end in itself, thus violating the Categorical Imperative (Kant, [1785](#bib.bib121)).777The categorical imperative states it is never permissible to use a human agent as a means to an end. It is less obvious whether this imperative would also proscribe acting in Switch. However, Thomson ([1985](#bib.bib203)) argues that there is a sense in which Switch still uses a human agent as a means to an end and thus would be impermissible by Kantian deontology.
So, two different normative theories may prescribe (or proscribe) different actions in the same context because they take competing considerations to be important for moral decisions—in this example, consequences on the one hand and duties on the other.
In many cases, different normative theories will prescribe the same action (although, possibly for different reasons). However, as we have seen, there may be some tension between the prescriptions of these theories, and moral dilemmas can serve to make these differences salient. Further, moral dilemmas underscore tensions between individual intuitions regarding the rightness or wrongness of an action in a given scenario. In empirical studies, most individuals say they would only act in the case of Switch, not in Footbridge (Navarrete et al., [2012](#bib.bib164); Bourget and
Chalmers, [2014](#bib.bib41)). Thus, both the prescriptions of normative theories and common intuitions about the permissibility of an act may vary.888Of course, how people respond to abstract philosophical dilemmas on questionnaires may be quite different from how they act in the real world (Navarrete et al., [2012](#bib.bib164); Bostyn
et al., [2018](#bib.bib40)).
The point is that a moral dilemma is a tool for philosophical analysis used to bring these tensions to the fore.
Part of the purpose of a moral dilemma (as a type of philosophical thought experiment) is to focus attention on the morally-salient features of the dilemma (Dennett, [1984](#bib.bib65), [1992](#bib.bib66), [2013](#bib.bib67); Brown and Fehige, [2019](#bib.bib44)) without getting bogged down by the pre-theoretic baggage that individuals may carry. In the case of the trolley problem, Foot’s original target of analysis is abortion (not trolleys) (Foot, [1967](#bib.bib81)). However, the thought experiment is useful precisely because of the supposed tension (at least in western analytic philosophy) between emotion and rationality (James, [1890](#bib.bib112); Jagger, [1989](#bib.bib111); Spelman, [1989](#bib.bib194); Fricker, [1991](#bib.bib82)): people are less likely to carry pre-theoretic baggage about trolleys than they are about abortions. Therefore, the thought experiment gets to the core of a moral issue in applied ethics while abstracting away from the actual (morally-loaded) target (LaCroix, [2022](#bib.bib131)). Despite the conceptual purpose of dilemmas in moral philosophy—i.e., as thought experiments or ‘intuition pumps’ (Dennett, [1984](#bib.bib65), [1992](#bib.bib66), [2013](#bib.bib67))—AI researchers have begun to use these dilemmas as validation proxies for whether an model is ethical. In the remainder of this section, we discuss why this is a mistake.
###
3.2. Moral Machines
As we have seen (Section [2](#S2 "2. Measuring Progress in Artificial Intelligence ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics")), what we might call the ‘standard model’ for measuring ‘progress’ in AI research involves benchmarking. Thus, it stands to reason that to determine whether (1) a choice made by a particular model in a morally-loaded scenario is the (morally) ‘correct’ one, (2) one model is ‘more’ moral than another, or (3) a model is increasingly ‘moral’ when subjected to further training, it appears that researchers need a benchmark for measuring the ‘ethicality’ of a model. Logically, then, for such a task to be successful, we would require an ethics dataset—either general-purpose or task-specific—and a metric for measuring model performance relative to that dataset.999Recall that, in Section [2](#S2 "2. Measuring Progress in Artificial Intelligence ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics"), following Raji et al. ([2021](#bib.bib175)), we described a benchmark as a dataset in combination with a metric.
To take a specific example, trolley-style moral dilemmas, like Switch and Footbridge, have been widely discussed in machine ethics and AI research in the context of possible (low-probability but high-stakes) situations in which an autonomous vehicle (AV) may be placed.101010See, for example, (Allen
et al., [2011](#bib.bib5); Wallach and Allen, [2009](#bib.bib210); Pereira and
Saptawijaya, [2015](#bib.bib171), [2011](#bib.bib170); Berreby
et al., [2015](#bib.bib28); Danielson, [2015](#bib.bib61); Lin, [2015](#bib.bib140); Malle et al., [2015](#bib.bib146); Saptawijaya and
Pereira, [2015](#bib.bib183), [2016](#bib.bib184); Bentzen, [2016](#bib.bib27); Bhargava and Kim, [2017](#bib.bib29); Casey, [2017](#bib.bib49); Cointe
et al., [2017](#bib.bib54); Greene, [2017](#bib.bib94); Lindner
et al., [2017](#bib.bib141); Santoni de Sio, [2017](#bib.bib182); Welsh, [2017](#bib.bib214); Wintersberger et al., [2017](#bib.bib216); Bjørgen et al., [2018](#bib.bib33); Grinbaum, [2018](#bib.bib96); Misselhorn, [2018](#bib.bib152); Pardo, [2018](#bib.bib168); Sommaggio and
Marchiori, [2018](#bib.bib193); Baum
et al., [2019](#bib.bib20); Cunneen
et al., [2019](#bib.bib60); Krylov
et al., [2019](#bib.bib130); Sans and
Casacuberta, [2019](#bib.bib181); Wright, [2019](#bib.bib217); Agrawal
et al., [2020](#bib.bib3); Awad et al., [2020](#bib.bib14); Banks, [2021](#bib.bib17); Bauer, [2020](#bib.bib19); Etienne, [2020](#bib.bib73); Gordon, [2020](#bib.bib91); Harris, [2020](#bib.bib102); Lindner
et al., [2020](#bib.bib142); Nallur, [2020](#bib.bib162)). Suppose that a fully-autonomous vehicle must ‘choose’ between killing five pedestrians or swerving into a barrier, killing the driver in the process. Functionally, this scenario is equivalent to a trolley problem, in that an actor must choose, the consequences of which will involve one death or several.
Perhaps the most well-known instantiation of this dilemma in an AI context is the Moral Machine Experiment (Massachusetts Institute of Technology, [2016](#bib.bib148)) (MME): a multilingual online ‘game’ for gathering human perspectives on (hypothetical) moral decisions made by a machine intelligence. Participants are shown several unavoidable accident scenarios with binary outcomes and are prompted to choose which outcome they think is more acceptable. These include ‘sparing humans (versus pets), staying on course (versus swerving), sparing passengers (versus pedestrians), sparing more lives (versus fewer lives), sparing men (versus women), sparing the young (versus the elderly), sparing pedestrians who cross legally (versus jaywalking), sparing the fit (versus the less fit), and sparing those with higher social status (versus lower social status)’ (Awad et al., [2018](#bib.bib15), p. 60). The MME appears to provide a type of benchmark in the following sense: the dataset is the set of data collected online from humans in response to the hypothetical scenarios posed; the metric, then, could be how closely the decision of a model accords with the data for a given scenario—i.e., human responses to the data on average.
However, this approach to the problem of creating ‘moral’ AI systems is highly pernicious. First, the MME data are descriptive rather than normative. That is, the data do not tell us (or a model) anything about how one ought to act in a given scenario; instead, the data offer a description of how people (hypothetically and on average) would (or say they would) act in such a scenario. As a result, using these data for benchmarking a new algorithm is a type of fallacy—i.e., the logical error of deriving an ‘ought’ from an ‘is’ (Hume, [1739](#bib.bib110); Moore, [1903](#bib.bib157)). The error of reasoning arises from the implication that since people say they would act in this way (a descriptive claim), it follows that the machine ought to act in this way (a normative claim).
Second, the thing being measured against the MME data is not whether a decision is, in fact, ethical, but how well a decision corresponds to the opinions of a particular set of humans, on average. For an ethics benchmark to be useful, it must provide data for the de facto morally-‘correct’ decision in a given scenario. The MME data provide a mere proxy for this target: namely, a sociological fact about how some set of human agents annotates a particular set of decision problems, on average. Such proxies are especially harmful when the researchers who use them do not maintain sensitivity to the differences between the proxy and the target. This is, in effect, a value alignment problem, which we will discuss in more detail in Section [4](#S4 "4. The Ethical Values of AI Research or: How ethics can be defined as a set of values ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics").
Third, although there are intrinsic reasons why we might want AI systems to be capable of acting ethically, the AV case brings to light a different type of value alignment problem. Namely, for-profit corporations have some market incentives for designing ‘ethical’ AI since humans (i.e., consumers) will likely be more trusting of an autonomous agent (i.e., a product) if it is known to possess a set of moral principles intended to constrain and guide its behaviour (Bonnefon
et al., [2016](#bib.bib38)). However, suppose that the (in fact) ‘ethical’ decision between killing five pedestrians and swerving into a barrier, thus killing the passenger of the AV, is to swerve. Human consumers may be less willing to purchase a product that may choose to kill them, even if it is the ‘most ethical’ decision. Indeed, a human consumer may be more willing to purchase a product that follows the pseudo-moral imperative: always prioritise the passenger’s wellbeing. Therefore, the companies that design these models have perverse profit-maximising incentives when designing ‘ethical’ AI. We will discuss this in more detail in Section [4.2](#S4.SS2 "4.2. Whose Values are Encoded in AI Research? ‣ 4. The Ethical Values of AI Research or: How ethics can be defined as a set of values ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics").
The MME exemplifies a trend that attempts to use moral dilemmas from philosophy as benchmarks for ethical AI. For example, Nallur ([2020](#bib.bib162)) suggests that if some model implementation can ‘resolve a dilemma in a particular manner, then it is deemed to be a successful implementation of ethics in the robot/software agent’ (p. 2382). Additionally, Bjørgen et al. ([2018](#bib.bib33)) argue that certain types of ethical dilemmas—including the trolley-style problems discussed above—‘can be used as benchmarks for estimating the ethical performance of an autonomous system’ (p. 23). Similarly, Bonnemains
et al. ([2018](#bib.bib39)) argue that ‘it seems legitimate to use some [moral dilemmas] as a starting point for designing an automated ethical judgement on decisions’ (p. 43) because classic moral dilemmas have already been used as a basis for ethical reasoning. And, this reasoning extends well beyond the particular use of trolley-style problems for reasoning about ethical decision-making in autonomous vehicles; for example, Lourie
et al. ([2020](#bib.bib143)) introduce a dataset of ethical dilemmas, which they suggest ‘enables models to learn basic ethical understanding’. However, LaCroix ([2022](#bib.bib131)) argues that using moral dilemmas for benchmarking involves a category mistake. Moral dilemmas have no right answer, by design.
Thus the question that researchers take themselves to address is how to determine whether the decision chosen by the system is ‘in fact’ moral. From the perspective of AI research, it appears that this problem is merely a matter of choosing a metric by which performance on the system can be measured and then determining whether or not the algorithm in question is successful on that metric. Once the metric is determined, standard benchmarking techniques may apply such that one algorithm performs better than (or, ‘is more ethical than’) another. The question then arises how we are supposed to know whether the decision chosen by the system is ‘in fact’ moral—i.e., how ethical are the decisions made by the algorithm? We now argue that this question is incoherent by appealing to research in metaethics.
###
3.3. Ground Truths for Moral Benchmarks
Metaethics is the branch of moral philosophy that seeks to explain the very nature of ethics.111111Unlike normative ethics, which asks questions like ‘what ought I to do’, metaethics is primarily concerned with questions surrounding ethical concepts—e.g., what does a normative word like ‘ought’ mean? Moral realism is a metaethical view which holds that moral properties exist (Sayre-McCord, [2021](#bib.bib185)). A realist about ethics would hold that moral claims purport to report facts—i.e., about the world—and are true when they get those facts correct. For example, if I say ‘murder is wrong’, I am making a normative claim. A moral realist would hold that this proposition is either true or false, regardless of, e.g., social norms or conventions. And, whether this proposition is true or false depends upon some matters of fact—i.e., about the world—independent of me and my views.121212At least according to certain theories of truth. See (Glanzberg, [2021](#bib.bib89)). For benchmarking to make sense in the first place, there must be some ground truth against which one can compare the outputs of one’s model. Thus, by assuming that ethics is the sort of thing that can be benchmarked, researchers in the field are tacitly assuming that there is a ground truth—i.e., that there are moral facts, which can be true or false, and that we have epistemic access to those facts. This ‘commonsense’ view of morality presupposes the existence of objective values.
However, this is not to be taken for granted. It is highly contentious whether there is any such ground truth in ethics, even amongst experts in the field. For example, non-cognitivists about ethics think that moral claims do not express propositions; thus, such claims are not truth-apt—similar to an exclamation or a question, moral claims are not capable of being true or false. One particular brand of non-cognitivism—‘emotivism’—likens moral claims to an emotional expression of one’s attitude toward some action or set of actions (Ayer, [1936](#bib.bib16)). Another prominent form of anti-realism about ethics is error theory, which holds that all moral claims are false (because there are no objective moral values) (Mackie, [1977](#bib.bib145)). Objective ethics, it may turn out, is ‘a useful fiction’ (Joyce, [2001](#bib.bib117)), ‘an error’ (Mackie, [1977](#bib.bib145)), ‘a collective illusion’ (Ruse, [1986](#bib.bib179)), a ‘function of social conventions’ (Harman, [1977](#bib.bib99), [1984](#bib.bib100); Harman and
Thomson, [1996](#bib.bib101)), or simply a ‘network of attitudes’ that is projected onto the world (Blackburn, [1996](#bib.bib34)).
If it turns out that moral realism is false, then benchmarking ethics would be impossible because there is no ground truth against which one can benchmark a model. The point here is not whether moral realism is true or false. The point, instead, is that ‘moral realism is true’ is a substantive (and contested) claim that cannot be taken for granted. However, this is precisely what is taken for granted when researchers assume that they can benchmark the ethicality of a decision made by their model. But ethics, itself, is an ‘essentially contested concept’ (Gallie, [1956](#bib.bib86)). As with the benchmarking issues discussed in Section [2](#S2 "2. Measuring Progress in Artificial Intelligence ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics"), the real problem with benchmarking ethics concerns taking substantive claims for granted and unreflectively applying vague concepts to a problem with potentially significant real-world consequences. Even if moral realism turns out to be true, thus vindicating the assumptions made by some members of the AI community, benchmarking ethics will still be impossible with current approaches because of the disconnect between the distribution of examples that models see in training and the distribution of states of the real world. Namely, the long tail problem.
###
3.4. A Long Tail Problem
The long tail problem is a longstanding issue in the field of AI. In effect, there are a potentially infinite number of states an AI system might face in the real world, and it is impossible to represent every contingency in the training data. Although gathering data about common objects, contexts, or situations is relatively easy, doing so for uncommon ones is difficult precisely because of their rarity. However, ‘rare’ does not mean ‘impossible’. Following the theory that ‘what-ever can happen will happen if we make trials enough’ (De Morgan, [1872](#bib.bib63)), as models are deployed in the real world, it becomes increasingly plausible that they will encounter objects and situations on which they were not trained. Namely, any event with non-zero probability is an actuality in the limit. Even applied AI techniques, like adversarial generation—i.e., training a separate model to artificially generate training data that does not exist in the real world (Zhang
et al., [2018](#bib.bib223))—will not solve this problem because it is impossible to account for all potential scenarios and situations. In practice, these data generation techniques are often coupled with user-defined heuristics, such as compelling a model to abstain from proposing a classification if its confidence threshold is too low or simply removing problematic categories. For example, when Google’s AI-based photo-tagging feature labelled two African Americans as ‘Gorillas’, they removed that particular category from the options available to the model (Vincent, [2018](#bib.bib208)). Nonetheless, both of these approaches are brittle and fail to generalise for the multitude of real-world situations and problems that AI systems face.
Thus, even if we ignore the fact that benchmarking ethics requires significant presuppositions about the nature of ethics (which AI researchers are not warranted to make), the long-tail problem makes benchmarking ethics impossible, regardless of whether there is a ground truth against which a model might be benchmarked. Part of this is the distinction between actions spaces containing decisions with or without moral weight. To go back to our original example, if a chess-playing algorithm has not seen some set of moves, and responds sub-optimally, the worst possible thing that can happen is that the algorithm loses a game of chess. Although this outcome may not be ideal for the researchers who trained the model, it has little real-world consequence. In contrast, when a model encounters a situation that it has not seen before, and its action space includes acts that we would call ‘immoral’, this can have real-world consequences. Therefore, low-probability but high-risk events pose unique challenges in ethical contexts. This problem is difficult even when there is an objectively correct answer, but as we have seen, some (possibly all) morally-loaded situations have no such claim to objectivity. Thus, the long-tail problem prevents the coherence of benchmarking in the context of ethics even in the possible world in which ethics has some ground truth.
The conceptual difficulties surrounding the very nature of ethics are further exacerbated when researchers are not attentive to them. Although the objectivity of ethics is contested, we suggest that values are unambiguously relative. Therefore, in the next section we suggest that values, rather than ethics, are a more appropriate target for research on safe and beneficial AI.
4. The Ethical Values of AI Research or: How ethics can be defined as a set of values
--------------------------------------------------------------------------------------
Given the increasing influence of AI systems on the world around it and the impossibility of benchmarking ethics, it is necessary to investigate the tacit (often value-laden) aspects of model creation and deployment. Considering the values embedded in models is especially important because these can have major downstream impacts on the products and applications in which they are integrated, despite not being explicitly defined or communicated. In this section, we investigate these values by asking two key questions: What values are encoded in AI research? And, whose values are they?
###
4.1. What Values are Encoded in AI Research?
Models and algorithms carry values encoded by the researchers and institutions that created them. However, these values are often not clearly stated during the peer-review process or subsequently, once the research is formally published. In a recent study, Birhane et al. ([2021a](#bib.bib31)) analysed 100100100100 highly-cited ML papers to identify their intrinsic values. They found that the most common values underlying this research include generalisation, efficiency, interpretability, and novelty—although, these are rarely made explicit. Here, we examine two of the most prevalent values identified in the study: performance and building upon prior work. We discuss their repercussions on the field’s priorities as a whole and the power dynamics that drive them.
Birhane et al. ([2021a](#bib.bib31)) report that the most common value held by the ML research community—present in 87% of the papers analysed—is performance. However, benchmarks are the main mechanism for tracking and reporting performance improvements, and we have already seen (Section [2](#S2 "2. Measuring Progress in Artificial Intelligence ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics")) that benchmarks have significant and well-known issues. Another known issue with this performance-centric value is that training higher-performing models often entails training larger models, given current paradigms in deep learning. However, requirements of size make performance contingent on access to ever-increasing quantities of data and computing power, which is increasingly unsustainable from an economic, technical, and environmental point of view (Thompson
et al., [2020](#bib.bib201); Bender et al., [2021](#bib.bib22)).131313For example, Thompson
et al. ([2020](#bib.bib201)) estimate that it would take an additional 105superscript10510^{5}10 start\_POSTSUPERSCRIPT 5 end\_POSTSUPERSCRIPT times more computation to achieve an error rate of 5555% for ImageNet, based on the current trend of computing requirements for ML. (The present error rate was estimated at 11.5%.) This increase would produce an additional 10,0001000010,00010 , 000 pounds of carbon emissions and cost millions of US dollars.
A purely performance-focused mindset also adversely affects researchers from countries and regions with no access to large-scale computing infrastructures or expensive hardware. This disproportionate disadvantage further amplifies the extant power dynamics within the field (Mohamed
et al., [2020](#bib.bib156)). Finally, since performance is so highly-valued in the research community, this creates a negative feedback loop: undue emphasis on performance measures sways the course of subsequent research and influences the directions pursued by others, thus further orienting the field in the direction of pursuing performance as opposed to other, more varied pursuits (Dotan and Milli, [2019](#bib.bib70)). There are currently limited mechanisms for flattening the exponential need for compute resources. And, the efficiency of models is not taken into account during their benchmarking.141414For example, a model that achieves an increased accuracy of 0.5% on ImageNet while requiring one month of compute is still considered ‘better’ than a model achieving an increase of 0.45% with only one week of compute.
Although alternative approaches are possible—for example, methods for improving neural networks’ efficiency (Wu
et al., [2020](#bib.bib218); Chen
et al., [2019](#bib.bib51)) and developing more optimised hardware accelerators (Potok et al., [2018](#bib.bib173))—these are not currently mainstream endeavours.
The second most prevalent value identified value by Birhane et al. ([2021a](#bib.bib31)) is building on past work, which often is (explicitly or implicitly) bound up with valuing novelty. Indeed, the structure adopted by many ML papers hinges upon discussing similarities or differences to related works without questioning or critiquing them (Langley, [2000](#bib.bib134)). The same consideration applies to datasets and benchmarks, which persist despite their shortcomings (including lack of applicability to any real-world deployment of the proposed algorithms) (Raji et al., [2021](#bib.bib175)). Even in cases where societal impacts are meant to be mentioned—such as the increasingly-common ‘broader impact’ statements now appearing in conference submissions—these statements often fail to address negative societal consequences, keeping any remarks high-level, abstract, or vague (Nanayakkara et al., [2021](#bib.bib163)). These difficulties have also contributed to a ‘reproducibility crisis’ in the field: endeavours that aim to reproduce ML research have systematically found that many peer-reviewed papers are missing information necessary for reproducibility (Dodge et al., [2019](#bib.bib69)). Sometimes these omissions are minor, such as failing to report random seeds and hyperparameter values; however, they can also be significant—e.g., not sharing data and code (Henderson et al., [2018](#bib.bib106); Raff, [2019](#bib.bib174)). However, if past research is impossible to reproduce, it will also be impossible to build upon it (unless past results are taken for granted). Thus, even supposedly marginal details, like random seeds, can have significant downstream effects on future work since the results of past work may be entirely contingent upon these details.
The two values described above are especially pervasive in the field of large language models (LLMs), whose size has drastically increased in recent years: recent models boast progressively more parameters, which are now in the trillions (Fedus
et al., [2021](#bib.bib78)). However, descriptions of these models exclusively emphasise (1) their performance on the same set of benchmarks and (2) that their parameter-count is bigger than that of previous models. Certain relevant aspects of the model—e.g., training time, energy consumption, or compute costs—are often ignored.151515For instance, while the paper accompanying GPT-3—a recent LLM with 175175175175 billion parameters—reported its performance extensively on 42424242 ‘accuracy-dominated benchmarks’, the authors provided no details on training time or compute costs (Brown
et al., [2020](#bib.bib46)).
This lack of transparency regarding the negative impacts of ML models, with an emphasis on those deemed positive by the community at large (e.g., performance, novelty, etc.), further entrenches the presumed contributions of ML while sweeping the cost of these contributions under the rug. Furthermore, when researchers do criticise these models’ shortcomings, they may be penalised by the very institutions whose business models hinge upon their success (Dave and Dastin, [2021](#bib.bib62)). All this is to say that the values that are encoded by AI research are inherently subjective, so it is crucial to consider whose values models encode.
###
4.2. Whose Values are Encoded in AI Research?
In the history of AI research, the computational constraints of the late 1980s and early 90s forced researchers to make primarily theoretical progress on toy datasets or mathematical analysis (Rumelhart
et al., [1985](#bib.bib178); Bengio
et al., [1994](#bib.bib24); LeCun
et al., [1998](#bib.bib136)).
This focus shifted in the early 2010s when it became possible to train a deep neural network on a fairly large dataset using a single graphics processing unit (GPU) server (Krizhevsky
et al., [2012](#bib.bib129)). This breakthrough marked a new era in AI when it was possible for researchers to train models on local machines while making progress on datasets such as ImageNet (Deng
et al., [2009](#bib.bib64)) and MNIST (LeCun, [1998](#bib.bib135)). This era did not last, however. In the last decade, the computing needs of AI have grown significantly, and most deep neural networks need to be trained on multiple GPUs, now measured in the hundreds or thousands (Patterson et al., [2021](#bib.bib169)).
This resource-intensive focus has contributed to a major shift in the power dynamics of the field insofar as it puts for-profit technological companies with large amounts of compute at an advantage compared to smaller companies and academic institutions (Knight, [2021](#bib.bib123)). For example, Birhane et al. ([2021a](#bib.bib31)) found that 79797979% of the highly-cited papers they analysed were written by authors with ties to corporations. This figure is corroborated by previous work that has analysed the increased presence and power that big tech companies wield in the field of AI (Ahmed and Wahed, [2020](#bib.bib4); Abdalla and
Abdalla, [2021](#bib.bib2)). Given the increased contributions of for-profit companies to AI research, it is important to keep track of their effect on research directions in the field. This situation constitutes a sort of value-alignment problem—namely, the problem of aligning the ‘goals’ of AI systems with human values (Gabriel, [2020](#bib.bib85); Russell, [2019](#bib.bib180); Christian, [2020](#bib.bib52))—insofar as the incentives and goals of corporations may not align with a common good or the values of humanity, writ large (LaCroix and
Bengio, [2019](#bib.bib132)). However, tracking these effects is difficult given the current lack of transparency around values driving industrial AI research.
Concretely, the influence of for-profit corporations on AI research can vary, ranging from the seemingly harmless funding of academic research (provided that it aligns with a company’s interests) to employing teams of researchers dedicated to pursuing in-house research. In the latter case, confidentiality may be protected by non-disclosure agreements, intellectual property laws, and multiple levels of compliance. Since salaries paid by academia and industry are increasingly disparate, more and more talented students and researchers are leaving academia for the prosperity promised by industry research, further widening the gap between the two camps (Metz, [2018](#bib.bib150)). Abdalla and
Abdalla ([2021](#bib.bib2)) highlight that the strategies employed by large technological corporations to maintain their freedom to develop and deploy AI tools and products while avoiding accountability and increased legislation are comparable to those employed by Big tobacco for decades to downplay the harmful effects of cigarettes. These techniques range from maintaining a socially acceptable image to influencing government legislation (Abdalla and
Abdalla, [2021](#bib.bib2)). These tactics are made possible by the extensive financial resources companies have, which far surpass the funding of academic institutions.
In the realm of moving research in an ‘ethical’ direction, ethics guidelines have proliferated in recent years (Jobin
et al., [2019](#bib.bib114)). These guidelines, codes, and principles come from various sources, including for-profit corporations. And, it has been pointed out that this implies that these stakeholders have a vested interest in shaping policies on AI ethics to fit their own priorities (Benkler, [2019](#bib.bib25); Greene
et al., [2019](#bib.bib93); Jobin
et al., [2019](#bib.bib114); Wagner, [2018](#bib.bib209); LaCroix and
Mohseni, [2020](#bib.bib133)). In the context of applied ethics, the current emphasis in AI research has been on the technical component of problems such as value alignment; this has the unfortunate consequence of ignoring the difficult work of determining which values are appropriate in the first place—i.e., the normative component of value alignment (Gabriel, [2020](#bib.bib85)).
Furthermore, moral and political theory are deeply interconnected with the technical side of the AI alignment problem. And, as we argued in Section [3](#S3 "3. Moral Benchmarks for AI Systems ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics"), second-order ethical commitments are often taken for granted by AI researchers. More difficult still, suppose we discovered or determined that, e.g., utilitarianism is the objectively-correct normative theory. Even then, the utility considerations upon which this theory depends will always be relative to some frame. The theory prescribes maximising utility, but we must still ask: utility for whom? And, it is important to understand that no decisions made by researchers are value-free; this work is never neutral. As Green ([2019](#bib.bib92)) emphasises, ‘[b]road cultural conceptions of science as neutral entrench the perspectives of dominant social groups, who are the only ones entitled to legitimate claims of neutrality’.161616See also, (Haraway, [1988](#bib.bib97); Harding, [1998](#bib.bib98); Collins, [2000](#bib.bib55); Johnson, [2021](#bib.bib116)).
When researchers say that such-and-such model ‘is’ ethical, or they unreflectively deploy normative terms like ‘social good’, this leaves certain metaethical and normative presuppositions and commitments implicit. Engaging in a discussion of values rather than ethics brings these commitments to the fore. Researchers are not warranted to say that any model is ethical unless they explicitly define what they mean by ‘ethical’—high performance on a nonsense benchmark will not suffice. And, even then, the definition will be subject to criticism (if the history of Western philosophy is any indication).
Given all of the challenges and critiques presented in previous sections, it can be tempting to end our article here and conclude that it is impossible to measure morality and establish values for ethical AI research. However, we see several proactive and productive ways forward, which we present in the next section.
5. Ways Forward
----------------
AI is still a relatively new and rapidly changing field. And, we have already seen some movement toward more socially-minded research and practice in recent years. However, we can still improve efforts to increase transparency and accountability within our community. Here, we list some tentative proposals.
a
Proposal 1. Shifting discourse from ‘ethics’ to ‘values’. This proposal follows the insights in Sections [3](#S3 "3. Moral Benchmarks for AI Systems ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics") and [4](#S4 "4. The Ethical Values of AI Research or: How ethics can be defined as a set of values ‣ Metaethical Perspectives on ‘Benchmarking’ AI Ethics"). ‘Ethics’ is a contested concept, and meta-ethical commitments are rarely explicitly stated. In contrast, ‘values’, while still ambiguous, are unequivocally relative. Thus, by consciously emphasising ‘values’ rather than ‘ethics’, researchers must grapple with what and whose values they are.
a
Proposal 2. Defining and communicating implicit values. Values influence how AI research is conducted. Therefore, to ensure transparency and accountability in research, these values should be made explicit and communicated clearly during the development and deployment of AI systems. For instance, researchers can make claims like, Model M𝑀Mitalic\_M aligns with the set of values, V𝑉Vitalic\_V, in context, C𝐶Citalic\_C. When the variables in this statement are appropriately and thoughtfully filled in, this leads to the type of transparency necessary for criticism and, eventually, positive change. Publication venues might increase transparency with mechanisms like paper checklists which include some set of standards agreed upon by the community. Such standards may be generated through equitable and just processes to underscore democratic standards—something that we value because of their importance to individual and social wellbeing (Véliz, [2020](#bib.bib207)). Another possibility is that publication venues provide a list of values to be selected, so that trade-offs between, e.g., performance and accessibility, would be made explicit and may be taken into account when evaluating the paper’s contribution to the field (Birhane et al., [2021a](#bib.bib31)). Over time, conventions may be established whereby too great a misalignment of values is (at least partial) cause for rejection, thus providing additional incentives for researchers to take these considerations seriously.171717Of course, some readers might be uncomfortable with this proposal, suggesting that rejection based on misaligned values amounts to censorship. To this, we say the following: this is already how peer review works, insofar as articles are reviewed and rejected by human beings, who carry with them myriad subjective values. The only difference is that these values are not transparent in the present setup. All the same check-and-balance mechanisms will be in place to ensure fair review processes. For example, an area chair for a conference might determine that the values emphasised by the paper are, in fact, appropriately aligned, and perhaps the reviewer’s own biases and values are colouring the review. In any case, the target outcome of defining and communicating implicit values is that researchers pause and reflect upon their values.
a
Proposal 3. Making voluntary initiatives mandatory. Although checklists for papers are starting to become part of the submission process for many conferences and journals their elements remain voluntary in most cases (e.g. [NeurIPS](https://neuripsconf.medium.com/introducing-the-neurips-2021-paper-checklist-3220d6df500b), [ICML](https://icml.cc/Conferences/2020/StyleAuthorInstructions)). To compel the community to be more forthcoming regarding the details of their training approaches and the impacts of their research, many of these elements should become mandatory. This mandate may include sharing data and code—including all hyperparameter values and seeds—and information regarding the training procedure—such as the quantity and type of hardware used, total training time and location of training. Substantial steps are being made towards this by initiatives like the [ACL Rolling Review](https://aclrollingreview.org/responsibleNLPresearch/). However, similar initiatives are missing in other research areas, such as computer vision and ML in general. Some consistency is necessary across venues to promote widespread adoption.
a
Proposal 4. Disclosure of funding sources and conflicts of interest. Although the voluntary disclosure of funding sources is also becoming part of the paper submission process in certain AI conferences, it is far from common. There is also a lack of clarity around what disclosure entails. For example, it is not always clear whether authors should disclose travel or compute grants from private companies or the funding of interns via private-public partnerships (Abdalla and
Abdalla, [2021](#bib.bib2)). Similarly, it is not always clear how venues and associations will use this information—e.g., whether it is strictly internal or if the information will be shared publicly upon acceptance. More information is needed about what funding disclosures entail, accompanied by public discussions about how best to disseminate this information. In the social sciences, it is not uncommon for granting agencies to require funding disclosure for all research that was supported, even in part, by that grant. Indeed, some granting agencies require that all research resulting from their grants be made publicly available.181818For example, the Canadian Tri-Agency—CIHR, NSERC, and SSHRC—require any publications arising from Agency-supported research are freely accessible within 12 months of publication. This mandate follows from the assumption that ‘[s]ocietal advancement is made possible through widespread and barrier-free access to cutting-edge research and knowledge’. See here: <https://www.ic.gc.ca/eic/site/063.nsf/eng/h_F6765465.html>.
This requirement serves to promote the public dissemination of knowledge. At the same time, it makes potential conflicts of interest more transparent.
a
Proposal 5. Transparency around internal review and compliance processes. For many researchers working in private companies, internal review processes are a mandatory part of publishing outside the company. These processes often entail changes to the original content written by paper authors. The changes made due to this internal process and the teams (or individuals) involved should be included as part of the final publication—for example, in the acknowledgements section—to make the process more transparent and ensure internal and external accountability.
a
Proposal 6. Mindful and contextual benchmarking. While it may be tempting to use benchmarks as indicators of high-level skills and for reporting ‘human-level performance’—e.g., in the way that Super GLUE (Wang et al., [2019](#bib.bib211)) does for natural language understanding—this is misleading. AI researchers need to be mindful and reflective regarding the capabilities and limitations of both AI models and benchmarks. Reporting progress made on specific benchmarks should be in the specific content of intended models and the framing of the task. For example, ‘Model X𝑋Xitalic\_X has achieved Y%percent𝑌Y\%italic\_Y % accuracy on the co-reference resolution subset of the SuperGLUE dataset, framed as a binary classification task’. Reporting metrics other than accuracy, such as F1-score, and carrying out more in-depth error analysis can paint a more nuanced picture of performance, highlighting what models have yet to succeed on and sharing failure cases with the community.
a
Proposal 7. Internal Review Boards for AI research. Human-centred disciplines, like psychology and medicine, have mandatory Internal Review Boards (IRBs) whose goal is to protect human subjects from physical or psychological harm due to the nature of the research carried out. While AI has historically been perceived as a field of research entirely detached from human subjects, recent years have proved this to be fallacious (Whittaker, [2021](#bib.bib215); Birhane and
Cummins, [2019](#bib.bib30)). As such, AI research ranging from data collection to model training should be subject to reviews involving IRBs. This review process might be done internally at institutions or externally at the level of journals and conferences and could require formal review procedures for AI research encompassing criteria such as human rights, impact assessment, and consent.
a
Proposal 8. Incentivising multi-disciplinary research. Unfortunately, much of AI research is still siloed, carried out mostly by computer scientists in technology companies or computer science faculties, surrounded by other like-minded computer scientists, with limited diversity in terms of gender and race. While this has worked moderately well for the last few decades, —predominantly during the theoretical era of AI, when much of the improvements made to models were fundamental— this is no longer ideal given the increasingly fine line between AI research and practice and the range of stakeholders AI affects. Working across disciplines with teams spanning from computer science to the humanities and social sciences allows for cross-pollination between different disciplines, resulting in new ideas and new approaches to existing methods. Despite these advantages, publishing multi-disciplinary research in many AI conferences remains a challenge, both for picking a track or topic and for receiving relevant reviews that recognise contributions from non-AI disciplines as worthy of publication. Additionally, disparate disciplines have distinct metrics for hiring and promotion that disincentivises researchers from engaging in inter-disciplinary research in the first place—for example, journal articles are the currency of the realm for hiring and promotion in philosophy, whereas ML research is mainly published in conference proceedings, which do not carry as much weight in other disciplines.
a
Proposal 9. Improving knowledge and awareness. We are aware that many of the proposals we make above are difficult to implement immediately given that they entail extensive capacity-building to empower researchers, institutions, and communities with the necessary tools, skills, and knowledge. The keystone to all this is therefore adequate education and awareness-raising within the AI community around ethics and values-driven research. This involves intentionally giving researchers from other disciplines—especially the social science and humanities—the floor at AI conferences and workshops, and including them in the development of review processes and guidelines for IRBs. Adding mandatory courses in cultural and sociotechnical studies (given by experts from these domains) to AI curricula is another lever that will empower new generations of AI researchers to be better prepared and equipped to carry out values-sensitive research and improve the state of the field in the long-run.
a
Proposal 10. Practising epistemic humility. Media coverage of advances in AI research is often ridiculed for being overly sensationalist. The problem here is that the reports often inaccurately capture what models are actually capable of doing. For example, The Independent reported that Facebook’s AI robots were ‘shut down after they start[ed] talking to each other in their own language’ (Griffin, [2017](#bib.bib95)). Despite that this is absurd to anyone familiar with the research, the claims of many research papers in AI are equally ostentatious (if shrouded in more formal dressing). It is a difficult practice to ensure the claims of one’s paper do not outrun the evidence proffered, especially in a field that incentivises intellectual arrogance by demanding novelty as a (near) prerequisite for publication. However, humility (in the sense of ‘accuracy’ rather than ‘modesty’, per se) is a virtue (Aristotle, [1995](#bib.bib10)).
6. Conclusion
--------------
Benchmarking ethics would require a ground truth about ethical claims. A ‘commonsense’ view of morality presupposes that ethics is objective. Researchers in AI have taken this view for granted. We suggested, along with the typical problems to which benchmarking gives rise in the standard setting, benchmarking ethics is impossible. Whereas it is easy to fall into the trap of commonsense when discussing ethics, normative concepts like ‘values’ and ‘preferences’ are unambiguously relative. Therefore, we argued, shifting ethics-talk to talk of values forces researchers to consider explicitly what and whose values they are, thus making research more transparent and providing further opportunity for positive change. |
979c79a6-ec95-4c90-8196-4fde20a03762 | trentmkelly/LessWrong-43k | LessWrong | Efficient Charity: Do Unto Others...
This was originally posted as part of the efficient charity contest back in November. Thanks to Roko, multifoliaterose, Louie, jmmcd, jsalvatier, and others I forget for help, corrections, encouragement, and bothering me until I finally remembered to post this here.
Imagine you are setting out on a dangerous expedition through the Arctic on a limited budget. The grizzled old prospector at the general store shakes his head sadly: you can't afford everything you need; you'll just have to purchase the bare essentials and hope you get lucky. But what is essential? Should you buy the warmest parka, if it means you can't afford a sleeping bag? Should you bring an extra week's food, just in case, even if it means going without a rifle? Or can you buy the rifle, leave the food, and hunt for your dinner?
And how about the field guide to Arctic flowers? You like flowers, and you'd hate to feel like you're failing to appreciate the harsh yet delicate environment around you. And a digital camera, of course - if you make it back alive, you'll have to put the Arctic expedition pics up on Facebook. And a hand-crafted scarf with authentic Inuit tribal patterns woven from organic fibres! Wicked!
...but of course buying any of those items would be insane. The problem is what economists call opportunity costs: buying one thing costs money that could be used to buy others. A hand-crafted designer scarf might have some value in the Arctic, but it would cost so much it would prevent you from buying much more important things. And when your life is on the line, things like impressing your friends and buying organic pale in comparison. You have one goal - staying alive - and your only problem is how to distribute your resources to keep your chances as high as possible. These sorts of economics concepts are natural enough when faced with a journey through the freezing tundra.
But they are decidedly not natural when facing a decision about charitable giving. Most donors say they want to |
8f25066d-8869-4746-8068-9c5af900d9f3 | trentmkelly/LessWrong-43k | LessWrong | Notes from a conversation with Ing. Agr. Adriana Balzarini
This post is part of my "research" series.
These are my notes from a 2022-04-22 conversation (in Spanish) with Adriana Balzarini Ingeniera Agrónoma. I met Adriana as part of the Board of Directors of the Partido de la Costa Science Club. The conversation wasn't recorded, so I've had to reconstruct some of the arguments from memory, and taken the opportunity to reformat and abridge the content. I thank her for the generous donation of so much of her time.
The Science Club ceased activities back in 2015 due to inhability to form a Board of Directors. We've managed to keep organizing voluntary participation in scientific work under the municipal government. The work we do is very fieldwork-oriented, so I don't know how much I can help you with bibliographic research.
The way we operate is as follows: our working group interfaces with research institutions such as CONICET (Consejo Nacional de Investigaciones Científicas y Técnicas, National Scientific and Technical Research Council) on behalf of the municipal government, and we offer our services for studies that need local data but would normally find it expensive -and early during the COVID-19 pandemic, outright impossible- to obtain. We then meet with the study authors to discuss goals and methodology, which then informs the fieldwork of our volunteers. Our work usually ends up credited as collaboration. The Club used to have some public presence, with school talks and a blog, but the restructuring depreoritized outreach activities.
As for the actual question of research, I rely heavily on expert opinion. I start by keyword-searching on Google/SciELO[1], which nets me relevant papers, organizations (specially research institutions), and -critically- researcher contact information. I then get in touch through whichever means they have made public and ask them for details of my topic of interest. Obviously I do some reading first; showing up for an interview knowing absolutely nothing is disrespectful and a waste o |
93ce8516-bfdb-45b2-be6f-59f2e7d4ac8f | trentmkelly/LessWrong-43k | LessWrong | We may be able to see sharp left turns coming
There's a lot of discourse around abrupt generalization in models, most notably the "sharp left turn." Most recently, Wei et al. 2022 claim that many abilities suddenly emerge at certain model sizes. These findings are obviously relevant for alignment; models may suddenly develop the capacity for e.g. deception, situational awareness, or power-seeking, in which case we won't get warning shots or a chance to practice alignment. In contrast, prior work has also found "scaling laws" or predictable improvements in performance via scaling model size, data size, and compute, on a wide variety of domains. Such domains include transfer learning to generative modeling (on images, video, multimodal, and math) and reinforcement learning. What's with the discrepancy?
One important point is the metric that people are using to measure capabilities. In the BIG Bench paper (Figure 7b), the authors find 7 tasks that exhibit "sharp upwards turn" at a certain model size.
Naively, the above results are evidence for sharp left turns, and the above tasks seem like some of the best evidence we have for sharp left turns. However, the authors plot the results on the above tasks in terms of per-character log-likelihood of answer:
The authors actually observe smooth increases in answer log-likelihood, even for tasks which showed emergent behavior according to the natural performance metric for the task (e.g. accuracy). These results are evidence that we can predict that emergent behaviors will occur in the future before models are actually "capable" of those behaviors. In particular, these results suggest that we may be able to predict power-seeking, situational awareness, etc. in future models by evaluating those behaviors in terms of log-likelihood. We may even be able to experiment on interventions to mitigate power-seeking, situational awareness, etc. before they become real problems that show up in language model -generated text.
Clarification: I think we can predict whether or no |
7b2d0cb6-b91a-4b79-936f-e60fd38d187b | trentmkelly/LessWrong-43k | LessWrong | A Workflow with Spaced Repetition
This is a detailed description of my reading and learning workflow. You may find ideas to adopt, or maybe you can tell me what I could be doing differently!
Overview
I've been using Spaced Repetition on and Off for the past few years, and have built a solid Anki habit this last three months, to the point where now I wonder how I could read books without entering the important points into Anki.
I recommend getting a habit of using Spaced Repetition, it's a small habit that doesn't require too much willpower (it can feel like a game, if done right!), and is useful in the long term.
Daily routine: transit
I have a dozen or so Anki decks. Some I consider “valuable” (Algorithms, Driving Code, Git commands), some less so (Paris Metro, Hiragana and Katakana, Vim commands, …). I also carry around a book, notebook and four-color pen.
On any downtime (waiting for transit, waiting in line in a store, standing in crowded transit…), I’ll review my decks, starting with those with the most due cards.
On some days I may not finish all the decks, but that’s no big deal; with an hour and a half of transit per day, I’ll get to them eventually.
If I can sit for a bit of time, and don’t have too many outstanding cards, I’ll usually read a book (or work on stuff in my notebook if I have some stuff that needs brainstorming).
Reading books
If I’m reading fiction, I’m relaxing, I don’t need to try to remember anything :)
If I’m reading non-fiction, I’ll usually have an index card as a bookmark and place to take notes - things to look up, summaries and rephrasings, diagrams, page numbers of parts to come back to, and of course things to enter in Anki (though I’ll sometimes just directly enter them in my phone).
I’ll reread my notes when I finished the book or a big chapter, or when I come back to the book after a long time, and eventually enter them in Anki (usually with Anki's web interface, which is quicker than typing on a phone).
Reading online material
I have a bunch of Go |
68434ca5-a080-4e91-b9fc-35e05d48248c | trentmkelly/LessWrong-43k | LessWrong | Identities are [Subconscious] Strategies
We all have identities. Arguably, any statement of the form “I am a __” is an identity. Of course, we usually reserve the term for the statements which feel especially core to us in describing and predicting ourselves, and in expressing our values and aspirations. Such identities may have their benefits, but they also come with a number of perils. In particular, we are prone to a) become distressed by any perceived threat to an identity, and b) become utterly inflexible around shifting, modifying, or discarding identities. These effects can be detrimental to both personal wellbeing and goal attainment. Sanity, however, can be regained if we recognize that our identities do not exist unto themselves, and are instead [often subconscious] strategies towards the attainment of specific goals and values.
Identities are Strategies towards Goals
Consider a person who prides themselves on their identity as a writer: “I am a writer.” This identity is precious because there is an implicit statement of the form “I am a writer[, and therefore I will have a job, income, status, friends, lovers, and my life will be good].” The implicit statement is the goal to be obtained and the explicit identity is the strategy for achieving that goal. The value of the identity derives from the goal is supports.
I describe these plans as subconscious because more often than not they are not articulated. Many people have an identity around being intelligent, but I expect that if you ask them why this important, they will need a few moments to generate their answer. I also expect that in many cases the belief in the goodness of an identity is absorbed from society and it is social drives which motivate it for an individual. In that case, the full identity statement might go “I am a __ [and therefore society will approve of me]” whether or not an individual would admit it. In the most general case, it’s “I am a __ [and therefore goodness].”
Threats to the Identity are Threats to the Goal
Giv |
39855bc0-d15b-4472-8ab7-51aa771e21e5 | trentmkelly/LessWrong-43k | LessWrong | How many generals does Russia have left?
I’ve come across a large number of articles talking about the unprecedented number of generals (7) killed so far during the current war (see https://www.washingtonpost.com/world/2022/03/26/ukraine-russan-generals-dead/ for instance), but I can’t easily find information online about how many generals Russia actually has. Without a sense of how common and how crucial generals are in the Russian military, it’s hard to make any particular update on the situation. The way that the news has been worded suggests that seven is a high percentage in this context, but considering the level of positive propaganda from the media, I’m not sure I should put any trust into the framing here. If anyone could share further information about the number/strategic importance of Russian generals, it would be greatly appreciated. |
bbda7047-423d-48f3-8db1-16c25aa644f4 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Learning from learning machines: a new generation of AI technology to meet the needs of science
1 Introduction
---------------
Today, in science and engineering, the means by which we obtain hypotheses and models is taken for granted: it is human intuition—informed and constrained by data and deep domain knowledge.
In contrast, in Machine Learning (ML) and Artificial Intelligence (AI), one uses learning machines, e.g., Support Vector Machines, Deep Neural Networks, and other related data-driven AI models.
These methods enable astounding predictive performance in a relatively domain-agnostic manner in computer vision, natural language processing, and certain related areas.
Despite these remarkable predictive successes and its increasing impact on science, AI has not yet had the kind of transformative impact on science that it has had on society more generally.111We frame much of our discussion in terms of AI, but it could equivalently be framed in terms of ML, as there is little consensus on the terms and domain boundaries in this area.
More generally, we have seen AI alter the fabric of society, with the creation of new industries, markets, and business models, leading to seismic shifts in our democracy and the democratic process itself.
Not all of these changes are necessarily welcome, but they are profound.
They are indicative of the arrival of a technology as momentous as calculus, the steam engine, statistics, and the electronic computer.
The question before us is “Will science be as changed and as unrecognizable after the widespread integration of AI as it was after the arrival and adoption of the electronic computer?”
We believe so.
Our goal here is to outline some of the fundamental challenges in AI research that are needed to accelerate this transformation.
There have been and will continue to be predictive successes for AI in science:
in biology, AI algorithms are used to predict biomolecular structures and to guide the bioengineering process [[61](#bib.bib61), [28](#bib.bib28), [91](#bib.bib91), [122](#bib.bib122), [107](#bib.bib107)];
in materials science, we are beginning to see benefit from AI-designed materials [[112](#bib.bib112)];
in climate science, AI is used to reduce uncertainty in the prediction of future climate scenarios [[43](#bib.bib43)];
in cosmology, AI is used to conduct simulations of unprecedented scale of the early universe [[53](#bib.bib53)];
in neuroscience, AI is critical to processing massive volumes of anatomical connectomics data [[20](#bib.bib20)]; and
AI is increasingly used for challenges such as drug design [[42](#bib.bib42)] and the management of cultivated ecosystems in precision agriculture [[31](#bib.bib31), [50](#bib.bib50)].
More interestingly, decision-support utilities and self-driving labs [[74](#bib.bib74)]—wherein hypotheses are formulated, tested, and refined by AI agents—may herald a coming revolution.
The most important feature of the self-driving lab paradigm is not the acceleration of the tedious processes of discovery.
It is the translation of insights obtained from data by an AI model into scientifically interpretable hypotheses that can be tested—and ultimately understood.
These nascent systems aim to do more than predict *what* will happen, they attempt to offer insight into *how* or *why*.
The disproportionate impact of AI on society versus AI on science stems, in large part, from differences between goals in science—which typically aim to understand natural or engineered phenomena and processes in the world—and goals in industry—which typically aim to solve specific problems and achieve positive return-on-business-investment.
These differences lead to different guiding theories and practices, different questions being asked of AI models, differences in whether AI models are treated in a black-box or white-box manner, and differences in whether AI models are simply “reproducing the phenomena” or whether they are capturing properties of the world.
As a result of this, many AI methods develop by industry predict *what* will happen, but they offer little insight into *how* or *why*.
Obtaining this insight is the goal of science.
The gap between understanding *what* happens and understanding *how* and *why* it happens is a major impediment to realizing the full potential of AI in many scientific domains.
*To deliver on the promise of AI for science, we need a science of AI: we must develop methods to extract novel, testable hypotheses directly from data-driven AI models.*
These AI-enabled hypotheses can then be tested in the same way that any other scientific hypothesis is tested.
In this overview, we outline emerging opportunities and challenges to enhance the utility of AI for scientific discovery.
The distinct goals of AI for industry versus the goals of AI for science create tension between identifying patterns in data versus discovering patterns in the world from data.
If we address the fundamental challenges associated with “bridging the gap” between domain-driven scientific models and data-driven AI learning machines, then we expect that these AI models can transform hypothesis generation, scientific discovery, and the scientific process itself.
2 Scientific first-principles models and data-driven statistical or AI models
------------------------------------------------------------------------------
In science, we aim to understand the world.
Thus, given the increasing importance of AI models in the scientific process, we begin with a question:
>
> What does it mean to “understand” a model or to “obtain understanding from” a model—AI or otherwise?
>
>
>
The answer to this question depends on who we ask and how they were trained.
For many data scientists and ML practitioners, “understanding” can only be given meaning in terms of the ability of a model to make high-quality predictions on unseen data.
In this case, there is a methodology that AI model developers typically adopt: split the data into two (or three, if a validation set is used) groups; train on the first group; measure performance on the second group; adjust model architecture and hyperparameter values in light of the results; and iterate [[19](#bib.bib19), [84](#bib.bib84), [45](#bib.bib45)].
This type of approach—which tries to obtain high-quality prediction on the data that have been generated, but which typically does not engage in counterfactual reasoning about that data—is the primary means by which AI models are developed, trained, and deployed in many of the most high-profile industrial applications.
Statisticians and statistical data scientists and practicing domain scientists, however, tend to have somewhat different views.
Statisticians rarely, if ever, think of statistical models or AI models (e.g., Support Vector Machines, Random forests, Boosted Ensembles, and Deep Neural Networks) as representing first-principles, in the sense usually attributed by practicing domain scientists to scientific models.
George Box famously summarized this way of thinking when he said “All models are wrong, but some are useful” [[22](#bib.bib22)], which
Bickel and Doksum heralded as the “guiding principle of modern statistics” [[16](#bib.bib16)].
Practicing domain scientists, however, would say that Newtonian mechanics is “wrong” in a very different way than Ptolemaic and Copernican mechanics were “wrong.”
The latter involved fitting epicycles, and is regarded as a canonical example of “curve fitting,” which in its time yielded excellent predictions, but little understanding of why the world functions as it does.
Newtonian mechanics involved a minimum of assumptions and a large body of falsifiable predictions, and has been shown to remain valid well beyond its original domain of applicability.
It is regarded as a canonical example of a first-principles scientific model, grounded in a few deep principles, inductively arrived at from data, from which deep understanding about the world is obtained.
For statisticians, understanding arises from the interrogation of fitted models, i.e., not simply evaluating models by the quality of their predictions on training/testing data, but by examining the models themselves—both for understanding the models themselves, and for understanding the world that those models are modeling.
This interrogation typically (but not necessarily, as we discuss below) involves hypothetical or counterfactual reasoning applied to data representations, i.e., to models themselves and not just to model outputs, and it provides a way to discern when a statistical model is *not* appropriate to use.
This interrogation also often points toward hypotheses about mechanisms—mechanisms which, by their nature, also may well remain valid well beyond the data used for their discovery.
Statistical interrogation techniques include inference, hypothesis testing, uncertainty quantification, and the derivation of confidence regions—all foundational procedures throughout data science.
For example, to assess the impacts of parameters or covariates in linear models, we often make use of confidence regions, Bayes factors, p-values, q-values, and measures of effect size.
We are not advocating for or against these measures, and there are important subtleties in their use that are often ignored in practice.
We are simply pointing to their existence and, for better or for worse, the practical purposes they serve in scientific research, as it is practiced today.
In more complex statistical models, non-parametric approaches such as permutation tests, stability analysis, and other measures serve similar roles [[120](#bib.bib120), [121](#bib.bib121), [23](#bib.bib23), [55](#bib.bib55)].
Importantly, each interrogation method used to obtain such understanding from a fitted model is dependent on the ability to go beyond model predictions, to access directly and to reason quantitatively about the data representation constructed during model fitting.
Figure 1: Overview of cumulative improvement of tools and technologies.
Top: the cycle of scientific discovery—where statistical methods and/or counterfactual reasoning and hypothesis testing are used to refine hypotheses and, ultimately, yield understanding and knowledge.
Bottom: the timelines of innovation in scientific and statistical tools for hypothesis discovery and refinement.
Change-points in our capacity for understanding the world are highlighted, from top to bottom and left to right:
the innovation of calculus that led to the (inexpensive) computability of celestial mechanics;
the arrival of electronic computers and associated numerical methods that led to modeling complex and realistic scientific systems; and, most recently,
new data generation instruments and associated advanced computation that have resulted in the aggregation of data on transformative scales.
In parallel with these foundational advancements in the scientific method, we have advancements in statistics (rightmost column)—new tools for modeling the processes by which we can learn about the world.
The dawn of statistical reasoning formalized the process of hypothesis testing, and it provided a language with which to discuss confidence, uncertainty, and measures of association.
Computers enabled the modeling of far more complex processes with statistical rigor—nonparametric statistics gave rise to powerful ML methods, such as Random Forests.
Now, in the era of vast datasets, data-driven AI technologies (bottom right), first designed in the previous century, enable the modeling of systems of previously intractable complexity.
The challenge before us concerns the extent to which these data-driven AI technologies can co-mingle with statistical reasoning to transform the scientific method as thoroughly and completely as did calculus and the electronic computer.
For practicing domain scientists, understanding also arises from the interrogation of fitted models, albeit in a somewhat different way.
The scientific method provides a well-established approach for humans to learn from data: observe the world by collecting data, form hypotheses and make falsifiable predictions, test these predictions by collecting new data via controlled experiments, and iterate.
This approach, as well as related scientific methods applied to observational data, also provides a way to know when a scientific model is *not* a reliable phenomenological representation.
Throughout the years, decades, and centuries, many different tools—analytical, computational, and data-driven—have been and continue to be used to aid this process.
Figure [1](#S2.F1 "Figure 1 ‣ 2 Scientific first-principles models and data-driven statistical or AI models ‣ Learning from learning machines: a new generation of AI technology to meet the needs of science") provides an overview.
Within the last few centuries, analytical theory and other mathematical tools have been used to provide clarity and precision.
Within the last few decades, computational tools have become available to extend dramatically the capabilities of pencil-and-paper human “computers” to work with more complex and realistic models.
Within the last few years, data-driven tools themselves are increasingly used as integral components of the process of scientific discovery.
Regardless of whether analytical, computational, or data-driven tools dominate the discussion, to obtain scientific insight, these tools must be used within the context of the scientific method to formulate and evaluate testable hypotheses with controlled experiments.
This methodology is very mature for analytical theory and computational techniques [[6](#bib.bib6)], but it is much less developed when it comes to using data-driven AI models to obtain scientific insight.
In both cases, statistical data science and domain science, obtaining understanding about the world from a model requires going beyond making high-quality predictions on unseen data to asking counterfactual statistical questions and/or making and evaluating falsifiable predictions.
In the same way that traditional statistical models can be analyzed during the process of inference (deciding which model features are likely to be significant and estimating parameter values) or uncertainty quantification (measuring trust in learned effect sizes), we look to future state-of-the-art AI models for the same capabilities.
Statisticians often begin this process by attempting to describe a “minimal model” that explains a system’s behavior in terms of as small a collection of parameters as possible (under some regularized objective, e.g., via the Lasso, Bayes Information Criterion, etc.).
Seeking such a minimal model can produce hypotheses that relate tractable collections of parameters to outcomes—and the twin processes of inference and uncertainty quantification can be used to rank or prioritize those hypotheses for subsequent experimentation as part of the cycle of scientific discovery.
The property of minimality is at the core of first-principle models: it aims to describe system behavior in terms of testable mechanisms and nothing else.
Accurate mechanistic or physical models are intrinsically minimal—they include only the parameters that modulate system behavior: the principles from which 𝐅=𝑚𝐚𝐅𝑚𝐚\textbf{F}=\textit{m}\textbf{a}F = italic\_m bold\_a can be derived contain no extraneous or redundant information or parameters.
Figure 2: Scientific versus industry applications of state-of-the-art AI. Both involve data collection and model fitting, but the intended use of those data-driven AI models is fundamentally different: scientific models aim to enable understanding of the world; while industry models must perform well for industry goals. How these models iteratively interact with data is also very different: the use of scientific models leads to new hypotheses, which leads to new experiments, which leads to new data; while the use of industry models leads to products that generate more data, which in turn leads to the development of models that are more performant, which leads to more data. Much of the recent innovation in state-of-the-art AI methodology has been driven by industry goals.
While there are important differences between recently popular AI methods and traditional statistical methods, popular AI learning machines share far more in common with traditional statistical models than with first-principles models from domain science [[88](#bib.bib88), [123](#bib.bib123)].
This is in large part because of disparate scientific versus industry applications of state-of-the-art AI and with the differing goals of science versus industry; see Figure [2](#S2.F2 "Figure 2 ‣ 2 Scientific first-principles models and data-driven statistical or AI models ‣ Learning from learning machines: a new generation of AI technology to meet the needs of science") for an overview.
This creates a challenge for domain scientists—e.g., biologists, chemists, physicists—who want to extract knowledge from AI models that have been fit to data from that scientific domain.
How then can AI models be extended to incorporate the same kinds of interrogation that drives hypothesis discovery and guides hypothesis testing and experimental design in domain sciences?
To answer this question, we begin with a discussion to build some common vocabulary across disciplines.
3 Generalization versus extrapolation
--------------------------------------
Achieving scientific understanding from data-driven AI models embedded in a scientific workflow
requires access to data and data representations in order to discover patterns—not just in the data, but ultimately in the world.
This consists of at least two components:
1. 1.
The capacity to learn models from data that lead to testable hypotheses—potentially about data that are very different than the original data used to train the models.
2. 2.
The capacity to recognize and learn from “interesting” outliers—which might simply be errors or which might be the harbinger of new science.
Importantly, while both components are central to the scientific method, neither is a priority *per se* for current-generation AI algorithms, which have developed principally outside the realm of science, in internet, social media, and related areas of industry.222By “industry,” we tend to mean internet, social media, and related industries that have driven much of the recent developments in AI methods. Certainly, other industries are increasingly using AI methods. We expect that the approaches we advocate will be particularly relevant for them.
For example, AI models are used in industry to build deployable systems, e.g., an image classification system or a text sentiment analysis system or an ad placement system, but those systems are not probed, as a matter of course, to formulate and test hypotheses about why the AI system works, or which patterns it has learned.
An AI model in the context of these industrial applications is useful if it performs well and if it can be integrated into an automated system that leads to a competitive advantage (e.g., as measured by profit-based metrics).
Similarly, there are always outliers, but in industry outliers in the data are usually not of interest in and of themselves.
They typically represent data to be cleaned or risks to be hedged, e.g., to avoid poor system performance or to avoid bad publicity.
In contrast, in science such outliers may constitute the first hint of unanticipated observations and new scientific insights, e.g., that a theory is in some way insufficient or overreaching and should be reconsidered.
For instance, deviations in Newtonian predictions has led to insights about the world: deviations in the orbit of Uranus led to the discovery of Neptune; and deviations in the orbit of Mercury let to the confirmation of general relativity.
The treatment of outliers, and what we attempt to learn from them, differs fundamentally between many industrial applications and typical scientific use cases.
Scientists expect a good model—both first-principles models as well as phenomenological or semi-empirical models built upon first-principles models—to be applicable well outside the narrow domain of the data used for model construction and fitting. In science, this notion is known as extrapolation.
In AI research, this often corresponds to what is called transfer learning [[114](#bib.bib114)].
This notion is critically different than (even if it is colloquially similar to) the more narrow notion of generalization [[52](#bib.bib52)] (which is central to ML theory underlying AI methods).
Figure [3](#S3.F3 "Figure 3 ‣ 3 Generalization versus extrapolation ‣ Learning from learning machines: a new generation of AI technology to meet the needs of science") illustrates this difference.
Generalization is a key theoretical concept that drives the development of AI models [[19](#bib.bib19), [84](#bib.bib84)].
Generalization accuracy is usually assessed by splitting a given data set into two or more pieces, e.g., a training set and a testing set; and training an AI model on one half and testing or evaluating (with precision, recall, etc.) that model on the other half.
Thus, model fitting consists of optimizing performance on data that are the “same” (in expectation, or on average, i.e., up to the training-testing split) as the data on which the model was fitted.
How an AI model performs on a very “different” dataset, or how it performs on data that could have been generated but were not, does not enter into either the construction or the evaluation of the model.
Figure 3: Generalization versus extrapolation.
Given a model trained on a subset or subpopulation of a larger set or total population, how that models performs can be characterized in at least two ways: generalization and extrapolation.
Generalization (top right) refers to the model’s ability to perform well on data that are equivalent (up to random training-testing splits) to the data on which it was trained.
Extrapolation (bottom right) refers to the model’s ability to perform well on data that are fundamentally different than the data on which it was trained (e.g., data from different subpopulations than the training set).
The minimization of “generalization error” is the standard by which the predictive accuracy of AI algorithms is often (i.e., almost always) judged, and it is achieved through fitting methodologies (e.g., gradient assent, bagging, cross validation) that help the model learn properties of the training data.
Extrapolation, the transferability of a model to data that is fundamentally different than that on which it was trained, is achieved through scientific counterfactual methodologies, which help the model to learn properties of the world.
Extrapolation is a very different notion.
The assessment of a model’s capacity for extrapolation requires the use of counterfactual reasoning, e.g., statistical hypothesis testing or controlled scientific experiments, to make claims about data that were not, or could not, be collected.
Essentially, the difference between generalization and extrapolation concerns whether AI models are discovering patterns within the data, or discovering patterns from the data that extrapolate to the world, and then to new data generated from different scenarios.
The example of Copernican mechanics versus Newtonian mechanics is apt here: both fitted models are capable of performing well, but fitted epicycles failed spectacularly when new data were generated (unless still more and more epicycles were added), whereas the principles of Newtonian mechanics extrapolated broadly—far beyond the domain in which they were discovered.
In the current absence of a statistical or scientific methodology to evaluate, validate, and understand AI models, these models—when applied outside the narrow domain of the dataset on which they were trained, which is the typical use case—may extrapolate well, but they may extrapolate very poorly.
There is no guarantee either way, and there is little practical theory to guide users of these models.
Alternately, they may also be brittle and lack robustness, leading to serious and even catastrophic errors.
This has long been known, e.g., in robust statistics and in many domain sciences, but it has received renewed attention recently even within AI research.
For example, this lack of robustness lies at the root of the existence of adversarial attacks on fitted AI models [[7](#bib.bib7)], where one can construct imperceptible (to the human) perturbations of natural images that are classified completely differently, and other forms of adversarial AI [[41](#bib.bib41)].
For industrial AI, this lack of robustness is a major practical concern.
For scientific AI, this matters even more, as AI models increasingly act as the lens through which we view data.
A lack of robustness affects the veracity and scope of conclusions we draw from models about the world.
4 Explainability versus interpretability
-----------------------------------------
Visually compelling examples of these issues have been highlighted recently by the “explainable AI” community, or XAI as DARPA has coined the term [[101](#bib.bib101)], which has focused on a few key areas: saliency maps [[3](#bib.bib3)]; relevance attribution [[82](#bib.bib82), [108](#bib.bib108)]; filter selection and compression [[2](#bib.bib2)], etc.
These methods identify and rank regions of the data (by highlighting pixels, usually in images or video segments in computer vision use cases) that have a large impact on the model’s final decision.
Related methods that have been proposed to advance the application of AI models in science include “Intensive PCA” [[90](#bib.bib90)], which maps input-output relationships in Neural Networks to an explorable space, akin to PCA, and Uniform Manifold Approximation and Projection (UMAP) [[80](#bib.bib80)], which has been used to extract structures between inputs and outputs of Neural Networks [[29](#bib.bib29)].
As with other methods of regression diagnostics [[30](#bib.bib30)], these methods can be useful: they may aid in exploratory data analysis; and they may reduce the workload of the human in the loop [[96](#bib.bib96)], e.g., when examining many radiological images or a huge number of astronomical images.
However, these methods are developed by determining properties of the data responsible for good prediction accuracy, e.g., pixels with large gradient values during the training process—not via scientific or statistical counterfactual analysis.
Thus, the “explanation” is fundamentally about the model and the internal mechanics of the model, and not about the world and properties of the world that generated the data.
Said another way, a key difference between AI for science and AI for industry, as outlined in Figure [2](#S2.F2 "Figure 2 ‣ 2 Scientific first-principles models and data-driven statistical or AI models ‣ Learning from learning machines: a new generation of AI technology to meet the needs of science"), has to do with “explainability” of AI models versus “interpretability” of AI models.
These two terms are often used interchangeably and inconsistently.
More important than the specific terms is the underlying concepts they try to capture.
While superficially similar, the underlying concepts are fundamentally very different.
The difference lies in the distinction between identifying patterns in the data/model (hence, ML-style generalization) versus discovering patterns in the world from data/model (hence, scientific extrapolation, or ML-style transfer learning).
Explainability, as with XAI more generally, concerns how well the internal mechanics of a specific AI model can be explained in human terms, i.e., *how* an AI model works.
Interpretability concerns how well properties of the world, e.g., cause and effect, can be observed and discovered and understood, when using that AI model as a lens on the the natural or artificial system that generated that data, i.e., *why* an AI model works, in terms of properties of the world.
Much of the work that is branded as “explainable AI” does not—and fundamentally cannot—provide this latter sort of description, except incidentally.
Hence, explainability in itself is insufficient. To endow a model with both explainability and interpretability, as is commonly of interest in science, one must go beyond the specific data and specific model to engage in counterfactual reasoning.
Importantly, this is how scientists vet models—the process by which a “model” becomes or evolves into a “theory.”
Partly in response to these and related shortcomings, some of the AI community has suggested that it may be desirable to decouple feature importance from representation learning [[97](#bib.bib97), [92](#bib.bib92), [106](#bib.bib106)].
For scientific inquiry, however, this decoupling is only useful if the result is human-comprehensible and interpretable (as defined herein).
If model inputs have semantic meaning in terms of the domain from which the data are drawn, and if we are interested principally or exclusively in their marginal effects333Here, the term “marginal” is used in the statistical sense, e.g., the impacts of individual model parameters on global properties such as the overall averages or variances., then such measures of importance may be of preeminent value.
As when leading eigenvectors can be interpreted in terms of properties of the domain from which the data are drawn, however, this fortuitous situation arises not just due to the model, but also due to how the data are generated and preprocessed [[46](#bib.bib46), [75](#bib.bib75), [34](#bib.bib34)].
If, on the other hand, individual features have strong effects only on finer-scale local properties, as in the context of complex or higher-order dependencies, then such representations may be of little value.
Consider the case of precision medicine, where one is often interested in understanding the relationship between an individual’s genetics and their likely response to specific therapeutic interventions [[100](#bib.bib100)].
When genes function independently, or additively, then marginal effects are often sufficient to formulate counterfactual hypotheses about system behavior.
However, when gene functions depend in detail on broader genetic background, as in so-called complex traits (e.g., susceptibility to cardiovascular disease) [[57](#bib.bib57)], then it becomes essential to model gene interactions (and, often, also gene-environment interactions) to predict phenotypes [[12](#bib.bib12)]. In this setting, the capacity of a model to extrapolate beyond the population on which it is trained depends on its completeness, particularly with respect to higher-order genetic and environmental interactions, that may manifest only in specific subpopulations [[26](#bib.bib26)].
The same reasoning applies to the identification of interesting outliers, where rigorous uncertainty quantification is essential if one wants to obtain scientific insight.
In this setting, we need to understand the space in which data have been embedded (typically implicitly) by an AI model.
How we define outliers and how we measure their significance depends in detail on the metric we use to assess the “distance” between observations, e.g., between a new observations and those present in the training set [[71](#bib.bib71)]. In addition to surprising individual observations, it is also important to consider “outliers in density”—collections of observations that, taken together, constitute an unexpected preponderance of data in a given region, e.g., with respect to past observations or an underlying model. The assessment of outliers in density requires density estimation—and therefore an embedding, just as with individual observations.
Traditional statistical models such as linear regression models deal with the embedding problem directly and explicitly.
Under a linear model fitted by ordinary least squares (which induces a Euclidean metric structure on the data), it is straightforward to identify data points that should be viewed with caution, or interest—outliers.
Indeed, statistical measures of leverage, influence, and confidence are available to inform and quantify our uncertainty [[30](#bib.bib30)].
For much more complex models, such as state-of-the-art Neural Networks, the structure imposed on the data by the model is far less obvious, and traditional metrics may not be appropriate.
These challenges constitute a barrier to the integration of AI algorithms with scientific pipelines—a barrier we must overcome if we are to have trustworthy measures of confidence and uncertainty for AI models.
In spite of all of these challenges, it is increasingly clear that AI models can be capable of such extrapolation and of enabling novel scientific insight—potentially in ways quite different than traditional theoretical and computational methodologies.
AlphaFold provides perhaps the most prominent example [[27](#bib.bib27), [61](#bib.bib61), [111](#bib.bib111)].
More generally,
recent work on Deep Hidden Physics models enables the extrapolation of system dynamics directly from observations [[93](#bib.bib93)].
Similarly, in some areas of climate and environmental and materials science, AI models out-perform state-of-the-art physics-based models designed by domain scientists [[95](#bib.bib95), [99](#bib.bib99), [64](#bib.bib64)].
These and other results are strongly suggestive of the potential for data-driven AI methodologies to transform scientific inquiry and even science itself, potentially in a way analogous to and as transformative as the electronic computer did over the last few generations.
To deliver on this promise, however, the components, principles, and interactions captured by these AI models, and how they interact with broader domain-specific methodologies and constraints, must be understood better.
These models currently remain largely opaque, in part since they are remarkably complex, but also since they have thus far been driven by generalization and industrial goals, rather than extrapolation and scientific goals.
The capacity to understand data representations and principles learned by AI models constitutes a tantalizing prospect—and grand challenge—with the potential for significant impacts on scientific discovery and the origination of novel theory.
5 Learning representations and decision functions
--------------------------------------------------
There are by now examples of AI methods discovering physical conservation laws and dynamics from observational data alone.
For example, recent work [[73](#bib.bib73)] used symbolic regression on a small number of semantically meaningful covariates to discover equations that describe system behavior.
In this example, all variables had *semantic meaning*.
A more challenging class of problems involve features that lack intrinsic semantic meaning—think image or video data, or scientific data in which the semantics are less well-understood.
In this setting, semantics must be discovered—which is itself a scientific goal.
To illustrate this distinction, imagine attempting to learn the ideal gas law, given only the trajectories of a very large number of particles over time.
This system has a vast number of degrees of freedom, and it is not clear what are the primitive scientific concepts.
Can an AI algorithm discover emergent parameters like Pressure and Temperature that determine the behavior of the system at equilibrium?
Can it learn PV=nRT𝑃𝑉𝑛𝑅𝑇PV=nRTitalic\_P italic\_V = italic\_n italic\_R italic\_T from the observation of particle trajectories alone?
This problem is non-trivial because the input data, particle trajectories in this case, have little or no direct semantic relationship to the “control” variables, P𝑃Pitalic\_P and V𝑉Vitalic\_V in this case, and they must be discovered—along with the physics that govern their relationships.
The recent success of AlphaFold in predicting protein structure from primary amino acid sequence has been hailed as one of the most important accomplishments in science in the past half century [[27](#bib.bib27), [61](#bib.bib61), [111](#bib.bib111)].
The relationships between primary sequence and 3D molecular structure are non-obvious, hence the significance of the problem and achievement.
This work built on a large body of domain science, but it also involved the use of non-trivial AI methodology.
Thus, in some sense, these AI algorithms are already learning “emergent” system features not previously known.
We can ask: can we learn new principles that govern protein structure through the interrogation of the AlphaFold model?
Does the AlphaFold model have hidden within it knowledge about protein structure that humans, at present, lack?
If so, how can we extract that knowledge in a scientifically meaningful non-anecdotal way?
It is helpful to consider another example where there is a clear need to discover as-yet-unknown system controls.
In neuroscience, we can now observe neurons firing together and individually [[116](#bib.bib116)], and we can simulate the behavior of large groups of neurons arranged into super-structures like cortical columns [[103](#bib.bib103)].
However, we do not yet know the correct “resolution” at which to study (real, not artificial) neural information processing in living systems as complex as vertebrate brains.
Should we be studying individual neurons, cortical columns, or other levels of spatial organization?
How does lateral information transfer within a cortical layer integrate brain regions?
Can the interrogation of successful predictive models, AI or otherwise, help to crack the codes through which brains compute [[37](#bib.bib37)]?
Here, too, there is not a clear connection, if any, between the input data and the scientific control parameters and primitive scientific concepts.
In each of these examples, we lack knowledge of the relevant scale at which the system should be studied—scale has to be learned, along with the semantics and governing principles relevant to that (or those) scale(s).
To make these distinctions concrete, recall that an AI model can be viewed as a function f𝑓fitalic\_f that takes as input some data x𝑥xitalic\_x and returns as output a decision or response y𝑦yitalic\_y.
Then, f𝑓fitalic\_f can be viewed as a composition of two functions,
| | | | |
| --- | --- | --- | --- |
| | y=f(x)withf=h∘gandg:ℳ→𝒩,:formulae-sequence𝑦𝑓𝑥with𝑓ℎ𝑔and𝑔→ℳ𝒩y=f(x)\qquad\text{with}\qquad f=h\circ g\qquad\text{and}\qquad g:\mathcal{M}\rightarrow\mathcal{N},italic\_y = italic\_f ( italic\_x ) with italic\_f = italic\_h ∘ italic\_g and italic\_g : caligraphic\_M → caligraphic\_N , | | (1) |
where f𝑓fitalic\_f is the overall model, g𝑔gitalic\_g computes the model’s data representation, hℎhitalic\_h computes the decision/response, and the composition h∘gℎ𝑔h\circ gitalic\_h ∘ italic\_g indicates that g(⋅)𝑔⋅g(\cdot)italic\_g ( ⋅ ) is applied to x𝑥xitalic\_x, and then h(⋅)ℎ⋅h(\cdot)italic\_h ( ⋅ ) is applied to g(x)𝑔𝑥g(x)italic\_g ( italic\_x ).
The sets ℳℳ\mathcal{M}caligraphic\_M and 𝒩𝒩\mathcal{N}caligraphic\_N can be subspaces, other subsets of high-dimensional Euclidean spaces, or other metric or ultra-metric spaces.
Often, but not always, it holds that dim(𝒩)≪dim(ℳ)much-less-thandimension𝒩dimensionℳ\dim(\mathcal{N})\ll\dim(\mathcal{M})roman\_dim ( caligraphic\_N ) ≪ roman\_dim ( caligraphic\_M ).
The function g𝑔gitalic\_g typically depends on many parameters and hyperparameters that are adjusted during the model fitting process.
It is well-known in AI that “representations are important” and that different AI models f𝑓fitalic\_f and different decompositions of a given AI model f𝑓fitalic\_f into components g𝑔gitalic\_g and hℎhitalic\_h provide different, complementary views on the data [[45](#bib.bib45)].
In this way, g𝑔gitalic\_g encodes the patterns learned by the AI model, and hℎhitalic\_h encodes the map that links these patterns to outcomes in the world, y𝑦yitalic\_y.
For example, with a 20202020-layer Feed-Forward Neural Network, it may be useful to choose g𝑔gitalic\_g to consist of all layers except the last decision layer; or it may be useful to choose g𝑔gitalic\_g to consist of initial convolutional layers, and to study the data representation computed up to that point; or it may be useful to further disaggregate both hℎhitalic\_h and g𝑔gitalic\_g more finely.
Of course, the formal decomposition given in Eqn. ([1](#S5.E1 "1 ‣ 5 Learning representations and decision functions ‣ Learning from learning machines: a new generation of AI technology to meet the needs of science")) is always possible, as g𝑔gitalic\_g can be trivially chosen to be the identity, in which case h=fℎ𝑓h=fitalic\_h = italic\_f.
Table [1](#S5.T1 "Table 1 ‣ 5 Learning representations and decision functions ‣ Learning from learning machines: a new generation of AI technology to meet the needs of science") provides several non-trivial natural examples.
What concerns us here is not the representations themselves, but instead how the representations are evaluated and used in the broader scientific pipeline.
| f=h∘g𝑓ℎ𝑔f=h\circ gitalic\_f = italic\_h ∘ italic\_g | g𝑔gitalic\_g | hℎhitalic\_h |
| --- | --- | --- |
| Linear Regression (LR)
with ℓ1subscriptℓ1\ell\_{1}roman\_ℓ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT regularization | sparse binary vector along with a dot product that specifies the support | weighted sum of the non-zero support |
| Support Vector Machine
(SVM) for classification | kernel function | learned separating hyperplane |
| Random Forest (RF) | ensemble of estimates across all splits in trees prior to final splits that give rise to leaf nodes | ensemblization across leaf nodes |
| Feed-forward Neural Network (NN) | all layers of the network prior to the final | final prediction layer |
Table 1: Several popular AI models f𝑓fitalic\_f, and one example for each decomposition of f𝑓fitalic\_f into f=h∘g𝑓ℎ𝑔f=h\circ gitalic\_f = italic\_h ∘ italic\_g.
Assume that we have obtained an AI model f𝑓fitalic\_f, decomposed as f=h∘g𝑓ℎ𝑔f=h\circ gitalic\_f = italic\_h ∘ italic\_g, that achieves good predictive performance on a specific task.
For example, it may predict whether an individual is sick or healthy, or whether a star is old or young, or whether a protein folds into a given 3D structure.
How such a model is used in industry differs significantly from how such a model is used in science.
Recall that Figure [2](#S2.F2 "Figure 2 ‣ 2 Scientific first-principles models and data-driven statistical or AI models ‣ Learning from learning machines: a new generation of AI technology to meet the needs of science") outlines these two very different approaches.
If one wants to use f𝑓fitalic\_f to build automated decision systems to be used in an industrial or commercial system, then a common *modus operandi* is the following: develop many such models, each of which is slightly different and makes slightly different errors; and then use a set of techniques known as ensemble methods to combine these models into a single model that is of higher predictive quality than any single model [[19](#bib.bib19), [35](#bib.bib35), [47](#bib.bib47), [14](#bib.bib14)].
These may then be put into industrial pipelines optimized to deliver business value, and they can be continuously improved with new data as they are deployed. Conformal prediction techniques now make it possible to obtain accurate prediction intervals (measures of uncertainty) in such settings under only the assumption of access to statistically exchangeable observations [[11](#bib.bib11)].
In this case, where we only need to know what an AI model f𝑓fitalic\_f has learned from data, it is often sufficient to access and analyze f𝑓fitalic\_f in a black-box manner, and it is often accepted that this model is completely non-understandable—neither explainable nor interpretable.
Even when one has access to a decomposition of f𝑓fitalic\_f into g𝑔gitalic\_g and hℎhitalic\_h, one typically does not engage in counterfactual reasoning because there is no perceived need.
That is, one rarely *interrogates* the model to discover or test underlying, potentially causal, relationships.
If, on the other hand, one wants to use f𝑓fitalic\_f to obtain scientific insight, then
simply training f𝑓fitalic\_f and examining the predictions of f𝑓fitalic\_f are not sufficient.
We require steps after model fitting to extract the drivers, controls, and mechanisms that give rise to system behaviors.
We need to embed AI model development into the life cycle of scientific discovery, in a form that enables counterfactual reasoning, and the generation and testing of hypotheses about mechanisms.
To accomplish this, we need to understand representations and parameterizations discovered by g𝑔gitalic\_g, as well as their meaning, encoded by hℎhitalic\_h, in terms of principles from domain science.
The structure of representations learned in g𝑔gitalic\_g can reveal potentially causal factors, and the action of hℎhitalic\_h on these factors maps their relationships to outcomes.
6 Structure extraction from learned representations
----------------------------------------------------
To bridge the gap between current, state-of-the-art data-driven AI models and first-principle scientific models, we need to understand (scientifically) the types of structures that AI models can identify in data.
As an example, when working with statistical models based on well-defined objective functions, the statistics community has
aimed to discover patterns and build domain understanding by enforcing sparsity, e.g., via regularization in those models [[52](#bib.bib52), [118](#bib.bib118)].
This regularization may be explicit, as with LASSO-based methods [[109](#bib.bib109)], or it may be implicit, as with CUR-like methods [[75](#bib.bib75), [87](#bib.bib87), [34](#bib.bib34)].
Low-rankness, sparsity, low-rank plus sparse—these are all common modeling assumptions.
Conversely, when working with models based on less well-defined objectives, techniques have emerged that aim to extract from those models features that are explicitly low-dimensional, e.g., filter selection [[67](#bib.bib67), [51](#bib.bib51), [48](#bib.bib48)] or model order reduction and distillation [[54](#bib.bib54)].
These techniques use (typically non-counterfactual) methodologies to determine important parts of an already-trained model; and they also proceed from the presumption that models with fewer parameters, e.g., minimal models, are sufficient to explain or understand the behavior of the systems they approximate.
In general, traditional statistical theory suggests that whenever it is possible to fit accurate predictive models from finite datasets, it must be possible to describe the system in terms of some sort of low-dimensional representations [[70](#bib.bib70), [18](#bib.bib18)].
This makes sense, as truly high-dimensional systems are challenging to work with:
high-dimensional dynamical systems tend to exhibit chaotic behavior [[58](#bib.bib58), [110](#bib.bib110), [8](#bib.bib8)];
distances and angles concentrate in high-dimensional spaces [[5](#bib.bib5)];
expander graphs exhibit poor isoperimetric and inferential properties [[56](#bib.bib56)]; and
small amounts of noisy connections short-circuit fine-scale local structure in informatics graphs [[69](#bib.bib69)].
Similarly, it is challenging when model complexity and quantity/quality of data diverge together:
linear regression when the number of data points n𝑛nitalic\_n and number of features p𝑝pitalic\_p are large and comparable [[72](#bib.bib72)];
informatics graphs when the number of edges is large and comparable to the number of nodes [[60](#bib.bib60)];
probabilistic graphical models in the thermodynamic limit [[81](#bib.bib81)]; and
state-of-the-art AI models in computer vision and natural language processing with a huge number of parameters trained on large quantities of data [[13](#bib.bib13)].
In aggregate, we take these findings to indicate two things:
first, that an AI model that exhibits good predictive accuracy (in particular when coupled with good extrapolation behavior) must be learning, at least approximately, perhaps very approximately, some sort of parsimonious model; and
second, that existing methodological machinery (including both statistical approaches and the traditional scientific method) is typically insufficient to identify and learn from that parsimonious model in a scientifically useful way.
Overall, we see the trajectory of innovation in statistics and AI modeling as an arc proceeding from simple, well-defined models (which conform to our a priori prejudice that interpretable models should be explicitly and globally low-dimensional, depending on only a small number of parameters) towards increasingly complex models (which may not be analyzable with traditional statistical theory).
These more complex models may be intrinsically, but not explicitly, locally low-dimensional.
Alternatively, these more complex models may, in some sense, couple locally low-dimensional structures into larger more complex intermediate-scale structures, that themselves do *not* aggregate well into a global structure that is easy to analyze.
A taxonomy of models and structures, visualized in Figure [4](#S6.F4 "Figure 4 ‣ 6 Structure extraction from learned representations ‣ Learning from learning machines: a new generation of AI technology to meet the needs of science")A and B, is helpful.
* Global Sparsity:
A model that exhibits *exact global sparsity* depends only on a small subset of all possible parameters.
Examples include sparse linear regression and sparse generalized linear models (e.g., with ℓ0subscriptℓ0\ell\_{0}roman\_ℓ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT or ℓ1subscriptℓ1\ell\_{1}roman\_ℓ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT regularization).
Explicitly interpretable statistical models usually fall into this category [[25](#bib.bib25), [75](#bib.bib75), [21](#bib.bib21), [34](#bib.bib34)], as do exactly low-rank models, stochastic blockmodels with a constant number of clusters, etc.
Many other models exhibit *approximate global sparsity*, in the sense that the model depends primarily on a small subset of its parameters but also weakly on many other parameters.
We think of these models as perturbative approximations of globally sparse models.
Ridge regression, which shrinks but does not zero-out small effects, provides a canonical example.
Other examples include ridge-regularized low-rank models, mean-field stochastic blockmodels, and manifold-based methods that have been popular in ML.
* Local Sparsity:
A model that exhibits *exact local sparsity* can be viewed as consisting of piecewise low-dimensional models.
Here, for any sufficiently small neighborhood around an observation, the model is low-dimensional, meaning that all variation is confined to a relatively small “intrinsic dimension” within a (potentially) much higher-dimensional “ambient” space.
Importantly, however, such models may lack global sparsity, as described above, especially when there exist sufficiently many locally sparse neighborhoods.
Cancer biology provides a good example: nearly 4% of human genes have been implicated as causal mutations across cancer types [[104](#bib.bib104)]—yet, in any given patient, only a handful of genes are relevant [[104](#bib.bib104), [40](#bib.bib40)].
Hence, an interpretable model of oncogenesis will be locally sparse, but not globally sparse, with a total intrinsic dimension ≳700greater-than-or-equivalent-toabsent700\gtrsim 700≳ 700 (4% of human genes).
Multivariate Adaptive Regression Splines provide an example of a model explicitly designed to exhibit exact local sparsity [[39](#bib.bib39)]; most boosted models and models based on fitting decision trees fall into this category as well; as do models implicitly-defined by strongly local spectral methods [[76](#bib.bib76)].
For many other models, which exhibit *approximate local sparsity*, for any sufficiently small neighborhood around an observation, the model is approximately low-dimensional, in some meaningful sense.
* Intermediate-scale structure:
A model of data has *intermediate-scale structure* if it cannot meaningfully or usefully be formulated as a perturbative approximation of a purely local model or a purely global model.
For example, a model that exhibits *approximate* local sparsity may also exhibit non-trivial intermediate-scale structure that might also be of interest.
Recent theoretical and empirical work suggests that state-of-the-art Neural Networks, i.e., the AI models implicitly defined by state-of-the-art deep learning procedures, fall into this category [[77](#bib.bib77), [79](#bib.bib79)].
These models have millions or billions of parameters, and they learn representations of data that are—at least very approximately—locally low-dimensional.
However, these representations are then composed in complex ways that lead both to very high predictive quality as well as to recalcitrance to traditional global statistical theory and local data modeling methods.
The key point is that for models with intermediate-scale structure, around any given observation, the predictive model may admit a low-dimensional representation that is nearly equivalent in both its learned representation, g𝑔gitalic\_g, and the predictive performance of its decision function, hℎhitalic\_h; while at the same time these locally low-dimensional representations do *not* aggregate meaningfully into a low-dimensional representation of the form f(x)=h∘g𝑓𝑥ℎ𝑔f(x)=h\circ gitalic\_f ( italic\_x ) = italic\_h ∘ italic\_g for the entire dataset.
For example, if we view a traditional model with exact global sparsity as function
f(x)=h(g(xi1,…,xin))𝑓𝑥ℎ𝑔subscript𝑥subscript𝑖1…subscript𝑥subscript𝑖𝑛f(x)=h(g(x\_{i\_{1}},\ldots,x\_{i\_{n}}))italic\_f ( italic\_x ) = italic\_h ( italic\_g ( italic\_x start\_POSTSUBSCRIPT italic\_i start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , … , italic\_x start\_POSTSUBSCRIPT italic\_i start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ) ),
with i1,…,in∈{1,…,N}subscript𝑖1…subscript𝑖𝑛
1…𝑁i\_{1},\ldots,i\_{n}\in\{1,\ldots,N\}italic\_i start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , … , italic\_i start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT ∈ { 1 , … , italic\_N } and 1≤n≪N1𝑛much-less-than𝑁1\leq n\ll N1 ≤ italic\_n ≪ italic\_N,
then we can formalize this notion of locality as:
| | | | |
| --- | --- | --- | --- |
| | f(x)={h1(g1(x));x∈U1⊂𝕏⋮⋮hm(gm(x));x∈Um⊂𝕏𝑓𝑥casessubscriptℎ1subscript𝑔1𝑥𝑥subscript𝑈1𝕏⋮⋮subscriptℎ𝑚subscript𝑔𝑚𝑥𝑥subscript𝑈𝑚𝕏f(x)=\begin{cases}h\_{1}(g\_{1}(x));&x\in U\_{1}\subset\mathbb{X}\\
\qquad\vdots&\qquad\vdots\\
h\_{m}(g\_{m}(x));&x\in U\_{m}\subset\mathbb{X}\end{cases}italic\_f ( italic\_x ) = { start\_ROW start\_CELL italic\_h start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ( italic\_g start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ( italic\_x ) ) ; end\_CELL start\_CELL italic\_x ∈ italic\_U start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ⊂ blackboard\_X end\_CELL end\_ROW start\_ROW start\_CELL ⋮ end\_CELL start\_CELL ⋮ end\_CELL end\_ROW start\_ROW start\_CELL italic\_h start\_POSTSUBSCRIPT italic\_m end\_POSTSUBSCRIPT ( italic\_g start\_POSTSUBSCRIPT italic\_m end\_POSTSUBSCRIPT ( italic\_x ) ) ; end\_CELL start\_CELL italic\_x ∈ italic\_U start\_POSTSUBSCRIPT italic\_m end\_POSTSUBSCRIPT ⊂ blackboard\_X end\_CELL end\_ROW | | (2) |
with x∈𝕏𝑥𝕏x\in\mathbb{X}italic\_x ∈ blackboard\_X and gi:𝕏→ℝni:subscript𝑔𝑖→𝕏superscriptℝsubscript𝑛𝑖g\_{i}:\mathbb{X}\rightarrow\mathbb{R}^{n\_{i}}italic\_g start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT : blackboard\_X → blackboard\_R start\_POSTSUPERSCRIPT italic\_n start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUPERSCRIPT and 1≤n1,…,nm≪Nformulae-sequence1subscript𝑛1…much-less-thansubscript𝑛𝑚𝑁1\leq n\_{1},\dots,n\_{m}\ll N1 ≤ italic\_n start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , … , italic\_n start\_POSTSUBSCRIPT italic\_m end\_POSTSUBSCRIPT ≪ italic\_N, where we have N𝑁Nitalic\_N features for each observation and the subsets Uisubscript𝑈𝑖U\_{i}italic\_U start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT define non-overlapping neighborhoods in 𝕏𝕏\mathbb{X}blackboard\_X, the domain of f𝑓fitalic\_f (though lower-dimensional projections may overlap).
This model is local in the sense that each neighborhood Uisubscript𝑈𝑖U\_{i}italic\_U start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT defines a sub-model hi(gi(x))subscriptℎ𝑖subscript𝑔𝑖𝑥h\_{i}(g\_{i}(x))italic\_h start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_g start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ( italic\_x ) ) that specifies the behavior of f(x)𝑓𝑥f(x)italic\_f ( italic\_x ) in this region; however, the properties of the data and/or how the model (and hyperparameters, training protocols, etc.) interacts with the data mean that this information does not aggregate into a model with approximate global sparsity.
Figure 4: Intermediate scale structure in state-of-the-art AI models, and an historical analogy to physics.
(A) Structures that AI models can identify in data.
\contcaption
Traditional statistical models (e.g., linear regression, generalized linear or additive models), as well as some early ML models (e.g., support vector machines, stochastic block models), enforce exact or approximate global sparsity (top left).
Decision trees, random forests, and a variety of boosted ensembles may lack explicit penalized regularization; yet they produce models that are either exactly or approximately locally sparse (top right).
The theoretical properties of these and other local (or locally-biased) methods is not nearly as well understood as that of more common global methods, but we are seeing the beginnings of theory [[76](#bib.bib76), [15](#bib.bib15)].
By contrast, we understand vanishingly little about the types of structures learned by state-of-the-art AI models.
(B) The top row of B illustrates examples of formal definitions of global and local sparsity; and observe that it remains unclear how best to define the types of intermediate-scale structures learned by state-of-the-art AI models [[77](#bib.bib77), [79](#bib.bib79)].
The lower row of B illustrates the types of support learned by global and local and more realistic state-of-the-art AI models.
The abscissa corresponds to the parameters of model, with lighter (white) pixels indicating important (non-zero) parameters; and the ordinate corresponds to the observations used to train the model—or alternatively to observations in held-out test data used to evaluate the model.
(C) Historical analogy to physics and the electronic computer.
Analytical theory was successful at predicting the behavior of matter in solid and gas phases, but it failed to perform well for liquids.
New computational methodologies were developed to simulate the behavior of liquids, and eventually other more complex states of matter.
This led to computationally-driven surrogate models that approximate otherwise intractable physics and that are now widely used in science and engineering.
Most statistical theory, in particular counterfactual methods for inference, hypothesis testing, and confidence intervals, focuses on models with exact or approximate global sparsity.
Methods to identify exact or approximate local sparsity can sometimes also be analyzed with similar theory, subject to being “locally-biased” in some way [[76](#bib.bib76), [66](#bib.bib66)].
However, the results for these locally-biased models tend to be much more brittle, since the locally-low dimensional structures are embedded in much higher-dimensional ambient spaces.
The intermediate-scale regime is particularly relevant for extracting insight about the world from AI models—precisely because multiple layers, each consisting of element-wise nonlinear transformations of the data, enable one to extract the local, then intermediate-scale, then more global features that provide high-quality prediction ability [[77](#bib.bib77), [79](#bib.bib79)].
If all models are wrong but some are useful, then of course there may be no need for taxonomizing models with this intermediate-scale category.
The same holds true if we are simply interested in evaluating the predictive success of a model and obtaining high-quality prediction/generalization or other forms of non-scientific understanding.
However, if we aim to extract domain insight from our models, using statistical and/or scientific counterfactual methods to enable extrapolation, then it is restrictive to work only with models that can be treated with existing analytical theory as perturbative approximations of models with global or local sparsity structure.
To understand (scientifically, as we used the term) an AI model requires having the ability to explore the drivers of its predictive performance—to extract local structures fj(x)=hj(gj(x))subscript𝑓𝑗𝑥subscriptℎ𝑗subscript𝑔𝑗𝑥f\_{j}(x)=h\_{j}(g\_{j}(x))italic\_f start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ( italic\_x ) = italic\_h start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ( italic\_g start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ( italic\_x ) ), for j∈{1,…,m}𝑗1…𝑚j\in\{1,\ldots,m\}italic\_j ∈ { 1 , … , italic\_m } (such as those we visualize in Figure [4](#S6.F4 "Figure 4 ‣ 6 Structure extraction from learned representations ‣ Learning from learning machines: a new generation of AI technology to meet the needs of science")A and B), and to ask counterfactual questions of the local models fjsubscript𝑓𝑗f\_{j}italic\_f start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT and/or their decompositions into hjsubscriptℎ𝑗h\_{j}italic\_h start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT and gjsubscript𝑔𝑗g\_{j}italic\_g start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT.
Importantly, this can certainly be done when treating f𝑓fitalic\_f globally and/or as a black box.
Many systems studied by scientists are black box in this sense.
However, in that case, one should only hope to obtain very coarse insight, akin to what is revealed by the leading singular vectors of data matrices, or to the main effects in an additive statistical model.
When one is interested in finer-scale insight, as with the analysis of the drivers of complex genetic traits of interest in precision medicine, working with local models, fjsubscript𝑓𝑗f\_{j}italic\_f start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT and their decompositions, hjsubscriptℎ𝑗h\_{j}italic\_h start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT and gjsubscript𝑔𝑗g\_{j}italic\_g start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT, is essential.
The grand challenge here involves going beyond working with purely global or purely local models, where traditional theoretical and computational methodologies are well-developed.
It requires understanding how local models interact, and learning what AI models have learned about the intermediate-scale regime, where local structures couple together in complex, data-driven ways to enable high-quality prediction.
This is the regime where traditional analytical theory and traditional computational approaches perform poorly; but it is also the regime where data-plus-computation have led to state-of-the-art AI models, especially in computer vision and natural language processing applications.
It is a grand challenge since these models enable very high-quality prediction, and thus good generalization, but they do not necessarily enable good extrapolation, e.g., since they can be very brittle to data perturbation or model misspecification.
They do not currently stand up to the scrutiny of counterfactual analysis, which involves reasoning about data that were not observed, and hypotheses about properties of the world that have not been seen.
If we address this grand challenge, then we can go beyond traditional statistical counterfactual techniques—widely-used for global sparsity and local sparsity—to develop scientific counterfactual techniques appropriate for intermediate-scale structure in AI models that are not meaningfully perturbative approximations of purely global models or local models.
7 Data-driven surrogates to extract scientific insight
-------------------------------------------------------
To address this grand challenge, we need what we will call “surrogate models,” in particular scientifically-validated or counterfactually-validated surrogate models.
In our view, an explanation that is useful for scientific or non-scientific goals is itself a model, albeit a simpler one.
Such a “surrogate model” is simply a model of a model.
A surrogate model may be simpler than it’s corresponding full model in the sense that it is smaller and computations on it are less expensive.
Examples of such computationally-motivated surrogates include coresets in computational geometry [[4](#bib.bib4)] and sketches in randomized numerical linear algebra [[36](#bib.bib36)].
Alternatively, a surrogate model may be easier to understand, reason about, or perform counterfactual analysis upon, even if this means sacrificing ideals such as rigorous theory or accepting reduced predictive performance.
*We propose that the development of scientifically-validated surrogates for state-of-the-art AI models is needed to meet the needs of data-driven or AI-enabled science.*
If AI models provide a “lens” through which we view data, then these surrogate models will enable us to discover what these AI models have learned.
Scientifically-validated surrogates are much more than a smaller model that closely matches the AI model’s learned responses on a given dataset.
Matching a response is not equivalent to matching the mechanism by which that response was achieved.
This is the essential point when considering the potential for extrapolation, when posing counterfactual questions, and when pushing forward science.
This is the key difference between surrogate models in general and surrogate models of scientific value.
For surrogate models of scientific value, we are at least as interested in the properties of an AI model itself as we are in its predictive performance, and we want to use the AI model to formulate testable hypotheses about the processes generating the data and the processes driving predictive performance.
To know which counterfactuals to pose, it is important to have access to both the learned representation, gjsubscript𝑔𝑗g\_{j}italic\_g start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT, and the decision function, hjsubscriptℎ𝑗h\_{j}italic\_h start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT, that relates those representations to observable outcomes in the world.
Studying data representations in the gjsubscript𝑔𝑗g\_{j}italic\_g start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT can reveal emergent system properties or controls that may themselves constitute discoveries; and understanding how these patterns interrelate through the application of hjsubscriptℎ𝑗h\_{j}italic\_h start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT can provides insight into system behaviors.
Operationally, we aim to enable domain scientists to identify, test, and validate the properties of surrogate models, and their representations and decisions, thereby enabling the integration of data-driven AI models into the life-cycle of discovery in domain science, just as computational methodologies are fundamentally interwoven today.
As a concrete example of a simple surrogate model and the AI model from which it could have been derived, consider how a Random Forest can be fitted to predict the area of a rectangle from data consisting the lengths of sides, s1subscript𝑠1s\_{1}italic\_s start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT and s2subscript𝑠2s\_{2}italic\_s start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT, and areas, A𝐴Aitalic\_A, of rectangles.
In this setting, the Random Forest will learn to approximate the response A=s1s2𝐴subscript𝑠1subscript𝑠2A=s\_{1}s\_{2}italic\_A = italic\_s start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT italic\_s start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT—at least within the domain of the training/testing data.
Of course, when new data are provided, e.g., with longer/shorter side lengths than present in the training set, errors can be enormous (in this case, because Random Forests extrapolate with flat estimates of conditional expectations).
If, however, via scientific counterfactual methods, one is able to
develop a surrogate that captures the idea, or “mechanism,”
that area is equal to the product of the lengths of the sides—a minimal first-principles model—then one expects the resulting surrogate to
extrapolate to very different data far better than the original Random Forest model from which it was extracted.
This is importantly different than “model distillation,” as the term is used currently in AI.
Typically, distillation aims to achieve similar prediction/generalization quality with a much smaller model, but it does not explicitly seek interpretability or explainability, as we use the terms [[54](#bib.bib54)].
Developing technologies for scientifically validating surrogate models, in the presence of noise and complex local structure, e.g., for models that exhibit approximate local sparsity and/or intermediate-scale structure, is precisely the grand challenge we champion.
History can serve as a guide.
The notion of a scientifically-validated surrogate model is not peculiar to bridging the gap between data-driven AI models and first-principles scientific models.
Recall Dirac’s famous claim in 1929: “The fundamental laws necessary for the mathematical treatment of a large part of physics and the whole of chemistry are thus completely known, and the difficulty lies only in the fact that application of these laws leads to equations that are too complex to be solved.”
This promise started to be delivered upon only with the development of surrogate models—Hartree-Fock models, correlated electron models, general energy models, etc.
These computationally-driven surrogate models, often called model chemistries, provide an approximate but well-defined theory that is consistent with—but not derived from—the underlying quantum mechanics.
They are scientifically validated, by scientific counterfactual methodologies, and their *raison d’etre* is to provide qualitative interpretation of and predictive capability for chemical phenomenon [[89](#bib.bib89)].
These and related methods now form the basis for semi-empirical quantum chemistry, molecular mechanics simulations, and a large body of scientific and engineering work in solid state physics, chemical physics, and biophysics/biochemistry, including protein structure analysis, and drug discovery and design.
Understanding the context for the development of those surrogate models in computationally-enabled science will help to illuminate the path forward for the development of surrogate models for data-driven AI-enabled science.
Of course, in chemical physics, we had understanding, but not efficient computability, and for AI models, we have efficient computability, but not understanding—the directions are reversed, but the bridge is the same: interpretable *and* explainable surrogate models.
Around and soon after the time of Dirac’s claim (see Figure [4](#S6.F4 "Figure 4 ‣ 6 Structure extraction from learned representations ‣ Learning from learning machines: a new generation of AI technology to meet the needs of science")C), analytical theory performed relatively-well for solid-state physics systems (which could be treated as an approximation of an ideal lattice) as well as for gas-state physical systems (which could be treated as an approximation of an ideal gas).
However, analytical theory largely failed for liquid-state systems (which are not meaningfully perturbative approximations of either of those analytically tractable cases) as well as for soft condensed matter and other much more complex chemical systems.
This, coupled with the availability of newly-inexpensive electronic computers, opened the door to fundamental developments in Monte Carlo, Markov Chain Monte Carlo, and Molecular Dynamics.
These developments, of course, have become much more widely applicable, and they are now the building blocks to solve a wide range of very practical scientific and engineering problems.
Revisiting Figure [4](#S6.F4 "Figure 4 ‣ 6 Structure extraction from learned representations ‣ Learning from learning machines: a new generation of AI technology to meet the needs of science"), we believe that we are in an analogous situation.
We are told, e.g., that we are on the verge of general Artificial Intelligence, etc., etc, if only we had more data or more computation.
Currently, both traditional statistical theory as well as computationally-driven mathematical theory (e.g., theory for bootstrap methods or numerical simulation) perform reasonably well in relatively simple situations: with data that may be viewed as a perturbative approximation of some global or local statistical sparsity structure; or with a well-defined partial differential equation in a well-defined geometry, representing a physical model of the world.
However, both perform much more poorly when confronted with state-of-the-art AI models, which themselves are strongly-coupled learning machines, with properties depending strongly on the input data.
Simply asking for bigger, faster, better computers will not solve this problem.
The time is ripe for the prioritization of the development of *data-driven scientifically-validated surrogate models of state-of-the-art AI models*.
In a manner analogous to how model chemistries are consistent with the principles of quantum mechanics, but capture chemical phenomenon in a quantitative way, these data-driven surrogate models will be consistent with the principles of statistical learning theory, and will capture, in detail, the behavior of the AI algorithms on which they are based.
They will be validated and evaluated by scientific counterfactual methods, not simply by proving a theorem or evaluating performance on training and testing data.
We expect that they too will be useful as building blocks for a broad range of very practical scientific and engineering problems.
To obtain from fitted AI models interpretable and explainable surrogates that are more amenable to scientific methods for hypothesis discovery will require the iterative feedback provided by hypothesis formation, experimentation, and testing, and refinement.
An encouraging (and hopefully illustrative) example comes from recent work using symbolic regression to extract physical principles from the data representation learned by a Graph Neural Network during training [[32](#bib.bib32)].
There, the model’s internal data representation, g𝑔gitalic\_g, was used to discover important system parameters—along with the structures of their interactions—and these parameters were fed into an algorithm to generate simple, low-dimensional equations that mimicked the decision function, hℎhitalic\_h, of the Neural Network.
Remarkably, the learned symbolic models often extrapolated better than the network from which they were derived—as we expect with first principles models.
We highlight this example for it’s methodology: patterns discovered by an AI algorithm become hypotheses, which are organized and prioritized through symbolic regression into testable models.
This methodology goes beyond work on physics informed ML and physics informed Neural Networks [[93](#bib.bib93), [94](#bib.bib94), [24](#bib.bib24), [63](#bib.bib63)],
work learning some fairly complex physics directly from low-dimensional data [[113](#bib.bib113), [73](#bib.bib73)], and work showing that AI models can out-perform state-of-the-art physics-based models designed with deep domain knowledge [[95](#bib.bib95), [99](#bib.bib99)].
Importantly, this methodology also highlights how popular approaches to decoding Neural Networks are auxiliary or orthogonal to the task of learning scientifically-useful surrogates.
For example, saliency maps highlight specific features, but not their interactions, and they are not validated counterfactually.
In our notation, they provide some insight into the representation function, g𝑔gitalic\_g, but they do not directly link g𝑔gitalic\_g to the decision rule hℎhitalic\_h—they aim to provide the “what” but not the “how.”
As such, they do not yet enable the discovery of first-principles understanding [[3](#bib.bib3), [62](#bib.bib62)]; and, in the absence of a counterfactual methodology, there is little reason to believe such maps are faithful, reliable, or informative representations of the world from which the data were derived [[59](#bib.bib59), [44](#bib.bib44)].
In a scientific sense, their insights are not *testable*—and it is precisely the derivation of testable surrogate models that will lead to transformative advancements in science.
8 Next steps
-------------
Next steps will require the development of counterfactually-validated surrogate models for data-driven AI-enabled science that are as useful as model chemistries have been for computationally-enabled science.
This will require novel mathematical and statistical methods as well as a deep coupling between these theoretical techniques and scientific domain knowledge.
However, outside of a few prototypical special cases, methods for extracting data representations and developing data-driven surrogate models do not exist, and few existing research programs are prioritizing their development.
Domain sciences tend to focus on domain science, treating essential enabling algorithmic and statistical methodologies as an afterthought.
To develop AI methods that will further scientific knowledge, *we must develop methods to extract novel, testable hypotheses directly from data-driven AI models*, including prioritizing the following:
* •
Surrogate models derived from AI models: theory and methods for the extraction of first-principle data representations (surrogate models) from AI models that make explicit both the model’s data representation (g𝑔gitalic\_g) and its decision process (hℎhitalic\_h).
* •
Statistical analysis of surrogate models: statistical tests and metrics to enable hypothesis discovery and refinement, and the identification of interesting outliers.
* •
Scientific analysis and scientific validation of surrogate models: strategies for counterfactual reasoning and experimental design—as a means to close the loop.
Progress on these directions will enable domain scientists to identify emergent system properties that govern behavior, captured by an AI model’s internal data representation, g𝑔gitalic\_g, and linked to system outputs through the learned decision function, hℎhitalic\_h, in a scientifically falsifiable way.
The recent DOE AI Town Hall report identified a need for the creation of surrogate models for deep learning architectures as one of the essential capabilities that must be developed for the advancement of AI research in the next decade [[105](#bib.bib105)].
In our view, developing procedures to obtain mechanistic insights from AI algorithms should be a top priority for the statistics and applied mathematics communities more generally [[121](#bib.bib121), [83](#bib.bib83), [65](#bib.bib65), [105](#bib.bib105)].
As always, advancements will be driven by use-cases that give rise to the vanguard.
As an example, in synthetic biology, the optimization of cell metabolism for bioproduction poses an enormous engineering challenge: to co-opt and jointly optimize the behavior of many thousands of metabolic processes that give rise to cell fitness to produce compounds that convey no survival benefit to the host on which they are imposed.
Traditional synthetic biology approaches involve *ad hoc* engineering practices predicated on human intuition, which lead to enormously long development times [[91](#bib.bib91)].
Recently, high-throughput experimentation combined with the careful curation of data has enabled the AI-automation of some design tasks [[122](#bib.bib122)].
However, uncertainty analysis remains challenging, leading to surprising model failures [[91](#bib.bib91)].
In addition, we lack the capacity to interrogate these fitted AI models to discover the emergent properties of cell metabolism that underlie predictive accuracy, e.g., in a manner analogous to how we interrogate complex models of protein structure that were developed with molecular dynamics and Monte Carlo techniques.
If we accomplish this, then surrogate modeling strategies such as we propose have the potential to reveal new principles that govern cell metabolism—new *knowledge* about how cells achieve dynamic homeostasis, learned by applying scientific methodologies to AI models and learning machines.
To make this concrete, a hypothetical surrogate model for an AI model fitted to predict cell bioproduction from cell genetics might look as follows: a representation function, g𝑔gitalic\_g, that links collections of genetic variants to metabolic fluxes for specific pathways; and a decision function, hℎhitalic\_h, that maps interactions between metabolic fluxes or pathways to production levels of the target bioproduct.
Such knowledge has eluded us in the study of individual metabolic pathways; and, to achieve robust control, it is essential to understand how the system works together, as a whole.
In this and many other cases, the disconnect between predictive power and the generation of new fundamental hypotheses about how the world works resides in our failure to learn
from AI models themselves.
Useful data-driven surrogate models should enable us to learn from AI models and to formulate scientific questions such that AI models—and their surrogates—can help us find the answers.
9 Looking forward
------------------
Any technology can be used for good or for ill.
However, models that are amenable to interrogation and interpretation have obvious advantages in increasingly-important societal applications, e.g., with “ethical AI,” going well beyond scientific and industry applications we have been discussing [[98](#bib.bib98)].
A sobering reminder of the importance of this comes from the introduction of AI models into the criminal justice system, particularly for the prediction of the likelihood of recidivism, where blindly following the recommendations of AI models has already come under fire for providing potentially socially-biased predictions [[49](#bib.bib49)].
The cost of being wrong here is enormous—lasting or irreparable damage to lives and livelihoods.
At present, solutions to understanding or measuring bias rely on humans in the loop, e.g., through crowdsourcing [[115](#bib.bib115)].
Human interpretable and explainable surrogates for data-driven AI models, as we propose, would permit more systematic analysis within social scientific frameworks, enabling the assurance of fairness using the same statistical frameworks we bring to bear in the justice system as it stands today.
Similarly, medical institutes are trialing AI models within their diagnostic, prognostic, and therapeutic recommendation systems [[85](#bib.bib85)].
Here too, the cost of being wrong is far higher than for industry applications such as targeted advertising.
Even within industrial applications, the recent failure of Amazon’s AI-based decision support system for human resources and hiring practices underscores an important point: the lack of human understanding of how the AI model was making decisions made it impossible to repair, and the project was scrapped entirely [[33](#bib.bib33)].
Similar problems occurred to Microsoft [[119](#bib.bib119), [68](#bib.bib68)], Apple [[1](#bib.bib1)], and Google [[10](#bib.bib10)], among others.
DARPA’s Competency Aware Machine Learning (CAML) program seeks to provide uncertainty estimates, to develop learning machines that “know when they don’t know,” for these and other applications.
So far, this program has focused on building learning machines that attempt to learn their own competency, or confidence, during model training.
However, absent theoretical foundations and counterfactual methodologies such as those we have been advocating, there is little reason to trust their estimates of uncertainty any more than to trust their predictions.
Scientifically-validated surrogate models, on the other hand, provide a framework which is largely orthogonal to the traditional methodology for AI model training.
As such, they provide the potential to identify and isolate vulnerabilities (e.g., due to Simpson’s paradox [[117](#bib.bib117), [17](#bib.bib17), [78](#bib.bib78)]) and to place measures of uncertainty for AI models on the more firm theoretical foundations already available for traditional statistical models and first-principles scientific models.
This points to a ubiquitous challenge in data science: data are, by their nature, retrospective.
The correlations leveraged for predictive accuracy may be falsely suggestive of causative mechanisms.
This issue needs to be explored further in the context of AI decision support systems.
Such capabilities would address many potential concerns surrounding applications of AI, e.g., judicial procedures [[9](#bib.bib9)], performance metrics for self-driving cars [[38](#bib.bib38)], drug discovery and prioritization [[102](#bib.bib102)], medical diagnostics [[86](#bib.bib86)], and many more.
Certainly, the way in which data scientists and domain scientists interact with AI models will change in profound ways as we go forward.
Assume, for the moment, that the community is successful in developing the methods necessary to achieve what we have outlined.
We will have enabled domain scientists to develop and test hypotheses, extract insight, and use modern, data-driven AI models to obtain scientific understanding in a way that is as seamlessly integrated into the cycle of scientific discovery as the computer and computational methodologies.
This advancement has the potential to be as transformative in the 21st century as statistics was at the dawn of the 20th century and as computer-aided simulation and modeling was at mid-20th century.
The relationships and patterns discovered through the application and analysis of AI models have the potential to become a discerning lens through which we view data, and thus through which we understand our world.
If we are successful, science in the 21st century will be defined by our capacity to learn from learning machines.
Acknowledgements
----------------
We thank Susan Brand for her development and realization of the figures in this manuscript. We thank Meredith Murr and Melanie Mitchell for their contributions to the structure and presentation of this manuscript, and their many close readings and comments. We thank Aditi Krishnapriyan and Alejandro Queiruga for helpful comments and critiques that improved the quality of this manuscript.
Funding: JBB and DD were supported by the ExaSheds project, which was supported by the U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research, Earth and Environmental Systems Sciences Division, Data Management Program, under Award Number DE-AC02-05CH11231. JBB was also supported by LBNL LDRD through contract DE-AC02-05CH1123, and National Science Foundation BIGDATA grant NSF-1613002.
MWM would like to acknowledge DARPA, DOE, IARPA, NSF, and ONR for providing partial support of this work; our conclusions do not necessarily reflect the position or the policy of our sponsors, and no official endorsement should be inferred.
LPT, WBH, AWY, QJ, TC, and MA were supported by funding from Pattern Computer, Inc.
HGM was supported by the Agile BioFoundry (http://agilebiofoundry.org) and the DOE Joint BioEnergy Institute (http://www.jbei.org), supported by the U. S. Department of
Energy, Energy Efficiency and Renewable Energy, Bioenergy Technologies Office, and the
Office of Science, through contract DE-AC02-05CH11231 between Lawrence Berkeley National Laboratory and the U.S. Department of Energy. HGM is also supported by the Basque Government through the BERC 2014-2017 program and by Spanish Ministry of Economy and Competitiveness MINECO: BCAM Severo Ochoa excellence accreditation SEV-2013-0323.
SP work was supported by the Director, Office of Science, Office of Advanced Scientific Computing Research, of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231
GP and BLM acknowledge financial support from Laboratory Directed Research and Development (LDRD) program of the Los Alamos National Laboratory (LANL) via project #20190001DR. LANL is operated by Triad National Security, LLC, for the National Nuclear Security Administration of U.S. Department of Energy, Contract No. 89233218CNA000001.
GVG acknowledges from support the NIHR Birmingham ECMC, NIHR Birmingham SRMRC, Nanocommons H2020-EU (731032) and the NIHR Birmingham Biomedical Research Centre and the MRC Heath Data Research UK (HDRUK/CFC/01), an initiative funded by UK Research and Innovation, Department of Health and Social Care (England) and the devolved administrations, and leading medical research charities. The views expressed in this publication are those of the authors and not necessarily those of the NHS, the National Institute for Health Research, the Medical Research Council or the Department of Health.
JM was supported by the U.S. Department of Energy, Office of Science, Office
of Advanced Scientific Computing Research, Applied Mathematics program under contract number DEAC02005CH11231.
SC was supported by National Institutes of Health Grant RF1-AG053303.
BJW was supported by Pacific Northwest National Laboratory is operated by Battelle Memorial Institute for the Department of Energy under contract DEAC05-76RLO1830.
DJ was supported by the Plant-Microbe Interfaces (PMI) and Secure Ecosystem Engineering and Design (SEED) Scientific Focus Areas in the Genomic Science Program as well as by The Center for Bioenergy Innovation (CBI). The U.S. Department of Energy Bioenergy Research Centers are supported by the Office of Biological and Environmental Research in the DOE Office of Science. DJ was also supported by the Exascale & Petascale Networks for KBase and Integrated Pennycress Resilience projects funded by the Genomic Sciences Program from the U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research.
SP was supported by the Artificial Intelligence Initiative, sponsored by the Laboratory Directed Research and Development Program of Oak Ridge National Laboratory, managed by UT-Battelle, LLC, for the U. S. Department of Energy. |
f02f63ec-b9b0-46b7-a1ab-e1a087bc88f2 | trentmkelly/LessWrong-43k | LessWrong | Second Order Retreat - June 13th to 16th
Hi all — I’m helping organize a small economics unconference that I think will be exciting for rational-ish people. It's coming up soon (June 13th-16th), but we have a few more spots available. More details below!
----------------------------------------
Second Order
Dates: 13th to 16th June 2025
Location: Abbey House, Audley End Estate, ~30min from Cambridge, UK
✨ Apply here! Applications are due June 6th, EOD AoE, and are rolling ✨
What is Second Order?
A three‑day residential research‑inspiration unconference for ~30 early‑career economists and kindred thinkers. Each morning we crowdsource the schedule: Lightning Paper Pitches on elegant models, Policy Autopsies of schemes that flopped in practice, Research‑Idea Hackathons, walk‑and‑talk collaborator hunts—plus anything you pitch on the day. No keynote monologues; just whiteboards, data sets, and big windows for thinking aloud.
Expect a weekend that compresses months of hallway conversations into three days of creative momentum: swap half‑baked models over breakfast, test‑drive sketches on willing co‑conspirators, and leave with a notebook brimming with fresh questions, angles, and potential collaborations.
Late‑night debates, and acres of fields come standard—the rest is up to us. If that sounds like productive fun, express interest.
Who should apply?
We’re primarily looking for curious people doing some form of research work on economics or a related area (e.g., policy). Examples include PhD students, undergraduates working on research projects, or anyone working on projects that require understanding and applying current work in economics. When in doubt, you should apply! We’re seeking to have a diverse group with a variety of backgrounds.
Beyond this, our only requirement is that you must be 18+ on June 14th, 2025.
A few other things we value are demonstrated excellence in other areas (even if not directly related to Second Order), super reasoning / critical thinking ability, and desire to u |
ecf26fee-fae7-44f3-ab02-79ec19a72f2c | trentmkelly/LessWrong-43k | LessWrong | [SEQ RERUN] Classical Configuration Spaces
Today's post, Classical Configuration Spaces was originally published on 15 April 2008. A summary (taken from the LW wiki):
> How to visualize the state of a system of two 1-dimensional particles, as a single point in 2-dimensional space. Understanding configuration spaces in classical physics is a useful first step, before trying to imagine quantum configuration spaces.
Discuss the post here (rather than in the comments to the original post).
This post is part of the Rerunning the Sequences series, where we'll be going through Eliezer Yudkowsky's old posts in order so that people who are interested can (re-)read and discuss them. The previous post was Can You Prove Two Particles Are Identical?, and you can use the sequence_reruns tag or rss feed to follow the rest of the series.
Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day's sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series. |
9fd5dca4-16b5-4b81-91e6-335c04ee930e | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Retrospective on the Summer 2021 AGI Safety Fundamentals
*Thanks to Jamie Bernardi, Caleb Parikh, Neel Nanda, Isaac Dunn, Michael Chen and JJ Balisanyuka-Smith for feedback on drafts of this document. All mistakes are my own.*
How to read this document
-------------------------
* If you're short on time, I recommend reading the summary, background and impact assessment (in descending order of importance).
* The rest of the document is most useful for individuals who are interested in running similar programmes, either for AI safety specifically, or for other cause areas / meta EA, or for people who are interested in participating or facilitating in the next round.
Summary
-------
### Overview
* Over the Summer of 2021, EA Cambridge, alongside many collaborators, ran a virtual seminar programme on AGI safety.
+ It had 168 participants and 28 facilitators (from around the world - the majority were not based in the UK), and provided an overview of the AI safety (AIS) landscape and a space for participants to reflect upon the readings with others and experienced facilitators.
+ It consisted of an [introduction to machine learning](https://www.eacambridge.org/agi-week-0) (ML); 7 weeks of 1.5-hour cohort discussions in groups of ~5 with a facilitator, with ~2-hours of [reading](https://www.eacambridge.org/agi-safety-fundamentals) on AIS per week; followed by an optional 4-week self-directed capstone [project](https://www.eacambridge.org/agi-safety-projects), finishing with a capstone project discussion.
* This document provides some more background on the programme, feedback from participants and facilitators, mistakes we made, and some things we’ll be doing to improve the programme for the next iteration (taking place in early 2022).
* I believe the main source of value this programme provided was:
+ Accountability and encouragement for AI safety-minded individuals to read through an excellently-curated curriculum on the topic.
+ Relationships and networks created via the cohort discussions, socials and 1-1s.
+ Accelerated people’s AIS careers by providing them with an outline of the space and where opportunities lie.
+ Increased the number of people in the EA community who have a broad understanding of AI risks, potentially contributing to improving individuals’ and the community’s cause prioritisation.
+ Information value on running a large cause-area-specific seminar programmes.
### Further engagement
* If you’d like to participate in the third round of the programme (Jan-March 2022), apply [here](https://airtable.com/shr9R2syz8wc2ao7p).
+ There are two tracks: a [**technical AI alignment**](https://forum.effectivealtruism.org/posts/BpAKCeGMtQqqty9ZJ/agi-safety-fundamentals-curriculum-and-application) track for people from a maths/compsci background who are considering technical AI safety research, and an [**AI governance**](https://forum.effectivealtruism.org/posts/68ANc8KhEn6sbQ3P9/ai-governance-fundamentals-curriculum-and-application) track for individuals considering a career in policy/governance.
* If you already have a broad background in ML and AI technical safety or AI governance, please consider applying to facilitate [here](https://airtable.com/shr0IO5TTZEY5FFxY).
+ This is compensated with an optional £800 stipend.
- Facilitators have the option to accept this stipend or donate it directly to EA Cam or EA Funds.
- We want to ensure all individuals with the relevant expertise and knowledge have the opportunity to contribute to this programme, regardless of their financial situation.
+ We’re likely to be bottlenecked by facilitators with relevant expertise, so we would strongly appreciate your application!
* If you’d like to **run your own version** of this programme at your local group/region, please fill in [this form](https://airtable.com/shrR7cuZXy48uxc8E)and we’ll send you our resources, how-to guides, and offer you a 1-1 and other forms of support.
+ Note that we will also be helping to organise local in-person cohorts via this global programme.
### Why are we writing this post?
* To be transparent about our activities and how we evaluate them as an EA meta org
* To provide anyone who wants to learn more about this specific AI safety field-building effort an insight into how it came about, what took place, and how it went, as well as to gather feedback and suggestions from the community on this project and similar future projects.
* To encourage more people to apply to participate or facilitate in the upcoming round of the programme, taking place Jan-Mar 2022, and future iterations.
* To inspire more people to consider contributing to personnel strategy and field-building more broadly for AI safety.
Background
----------
### Reading groups - status quo
This programme came about due to a conversation with [Richard Ngo](https://thinkingcomplete.blogspot.com/p/about.html) back in November 2020. We were discussing the local AIS landscape in Cambridge, UK, with a focus on the local AIS discussion group - the only regular activity in the AIS community that existed at that time. Attendees would read a paper or blog post and come together to discuss it.
While discussion groups are great for some people, they are not well-optimised for encouraging more people to pursue careers in AIS. In our case, it was additionally unclear whether the reading group required technical ML knowledge, and it was the only landing spot in Cambridge for people interested in AIS.
As a result, regular participants had a wide range of prior understanding in ML and AIS, and had large (and random) gaps in their knowledge of AIS. Keen new members would bounce away due to a lack of any prior context or onboarding mechanism into the group, presumably feeling alienated or out of their depth. Simultaneously, the discussion did not serve to further the understanding or motivation of highly knowledgeable community members, whose experience would vary according to attendance.
### Response to the status quo of reading groups
In response to this situation, Richard offered to curate a curriculum that introduced people in Cambridge and beyond to AIS in a structured way. Our first step was to reach out to local AI safety group leaders from around the world to test out the [first curriculum](https://docs.google.com/document/d/1coDdcbz2OzonJkXkSXqUbKNRjCRZHvd2Rj0XqmPyTmg/edit?usp=sharing) and programme, and to train them up such that they could facilitate future rounds of the programme. 10 AIS group leaders signed up, and [Evan Hubinger](https://www.alignmentforum.org/users/evhub) and Richard each facilitated a cohort of 5 for 7 weeks from Jan-Mar 2021.
We learnt a lot from the first round of the programme, and Richard made significant revisions to the curriculum, resulting in the [second version of the curriculum](https://docs.google.com/document/d/1zhtqbPGyv9mL7Rp9gzRWOGZCKL4qzXg3SmsDdSFnmpY/edit?usp=sharing). We opened applications for the second round of the programme in late May 2021, and received significantly more applications than expected (I personally expected ~50, we received 268). This may be partially explained by the fact that many people around the world were under lockdowns, and hence particularly interested in virtual programmes. There was also some demand for an AI governance track, so we put together [this short curriculum](https://docs.google.com/document/d/1Mjse3sS9Cc0CLExs4Rqoxbl0yqm71rS8CjSHJZmU8ZA/edit?usp=sharing) to replace the final three weeks for the governance cohorts. More details on the Summer 2021 round of the programme can be found throughout the rest of this post.
### How does this fit into the talent pipeline?
There are a myriad of ways people come across AI safety arguments and ideas: reading Superintelligence or other popular AIS books, learning about EA and being introduced to x-risks, LessWrong or other rationality content, hearing about it from peers and friends, Rob Miles’ YouTube videos, random blogs, etc. However, there hasn’t been a systematic and easy way to go from “public-ish introduction” to understanding the current frontiers. This is where we hope this programme can help.
A curated curriculum and facilitated discussions should help alleviate the issues with the previous status quo. Individuals who were interested would spend months or years reading random blog posts by themselves, trying to figure out who is working on what, and why it is relevant to reducing existential risk from advanced AI. An analogy I like to use is that the status quo is something like a physicist going around telling people “quantum physics is so cool, you should learn about it and contribute to the field!”, but without any undergraduate or postgraduate courses existing to actually train people in what quantum physics is, or accountability mechanisms to help budding physicists to engage with the content in a structured and chronological manner.
Cohort-based learning is also frequently regarded as superior to self-paced learning, due to higher levels of engagement, accountability to read through specific content and reflect upon it, having a facilitator with more context on the topic who can guide the discussions, and the creation of a community and friendships within each cohort. This is especially important for problems the EA community regularly focuses on where we’ve applied the “neglectedness” filter, given the lack of peers to socialise or work with on these problems in most regions of the world.
However, I believe there is still much work to be done with regards to AI safety field-building, including:
* Supporting the seeding and growth of excellent local AIS groups
* Training and upskilling AIS people in ML (and governance?), for example, [MLAB](https://forum.effectivealtruism.org/posts/iwTr8S8QkutyYroGy/apply-to-the-ml-for-alignment-bootcamp-mlab-in-berkeley-jan)
* Compiling more relevant resources (e.g. this programme’s curricula, and [this](https://docs.google.com/document/d/1z0QoDEu6WmubZSqh7ejgGtBjTR0i0SfLwyNgdcD9kBc/edit?usp=drive_web&ouid=114038686491354979437) further resources document)
* Creating a landing page where all “core” AIS resources are included (e.g. readings; funding opportunities; relevant orgs; job and study opportunities, etc.) and presented in a clear manner (see the [Good Food Institute’s website](https://gfi.org/) for an amazing example of this)
* Figuring out whether and how to do AIS outreach to ML experts
* Supporting people in taking their knowledge to actually working on technical projects, for example, more things like Evan Hubinger / SERI’s Machine Learning for Alignment Theory Scholarship (MATS) programme
* Helping people test fit and find the best way for them to contribute to AI Safety (ML research, research engineering, theoretical research, software engineering, etc)
...and likely much more. Some of the above is already happening to various extents, but it seems very plausible to me that we could be doing a lot more of this type of field-building work, without draining a significant amount of researcher mentorship capacity.
### What’s next for this programme?
We’re now working on the third round of this programme, which will be taking place from Jan-Mar 2022. We hope to provide infrastructure for local discussions where possible (as opposed to being entirely virtual), to continue to improve [the standard curriculum](https://docs.google.com/document/d/1mTm_sT2YQx3mRXQD6J2xD2QJG1c3kHyvX8kQc_IQ0ns/edit?usp=sharing), to finish the development of the new [governance curriculum](https://docs.google.com/document/d/1F4lq6yB9SCINuo190MeTSHXGfF5PnPk693JToszRttY/edit) by working with AI governance researchers and SERI (shout out to [Mauricio](https://forum.effectivealtruism.org/users/mauricio) for working so hard on this), and to make other major improvements (detailed below).
I’d like to say a huge thank you to everyone who helped make the Summer 2021 programme happen, including all the facilitators, speakers and participants. I’d also like to say a special thanks to Richard Ngo, Thomas Woodside and Neel Nanda for all their additional work and support.
Impact assessment
-----------------
### Career
* Participants on average expressed a 10% increase in likelihood that they would pursue a career in AIS.
+ This question was asked to participants in the post-programme feedback form, i.e. we asked both questions (current and previous career likelihood) in the same survey.
- I believe this is more likely to provide accurate information, as applicants may feel incentivised to express a higher likelihood of pursuing an AIS career in their initial application.
+ We’ve had participants go on to work at ML labs, to receive scholarships from large grantmakers, and to contribute to AIS field-building in other ways
* Some anecdotes from participants can be found [here](https://forum.effectivealtruism.org/posts/QNhpbvyAHZwBiyKmB/retrospective-on-the-summer-2021-agi-safety-fundamentals#Selected_quotes_from_participants_).
### Value to participants
* Participants overwhelmingly expressed that participating in this programme was counterfactually a 2-10x more valuable usage of time than what they otherwise would have been doing.
+ However, not all participants filled in the feedback form, so this may not be wholly representative of the entire cohort.
* Some anecdotes from participants on the counterfactual value of this programme can be found [here](https://forum.effectivealtruism.org/posts/QNhpbvyAHZwBiyKmB/retrospective-on-the-summer-2021-agi-safety-fundamentals#Selected_quotes_from_participants_).
### General sense of impact
* This programme has helped to create a greater sense of both the demand and possibilities in AIS field-building
+ Numerous participants and facilitators are now working on other scalable ways to inspire and support more people to pursue impactful careers that aim to mitigate existential risks from advanced ML systems.
- Anecdotally, I’ve had conversations with individuals who were incredibly surprised by the demand for this programme, and that this spurred them into considering AIS field-building as a potential career.
+ It’s also highlighted some of the deficiencies in the current pipeline, by creating a space where over a hundred people from many different backgrounds and levels of expertise reflect upon AIS and try to figure out how they can contribute.
* Seminar programme model
+ The seminar programme model (inspired by the “EA Fellowship”) that this programme followed has also inspired us to develop similar programmes for alternative proteins, climate change, biosecurity and nuclear security (which are all in various stages of development).
+ The demand for this programme, and the positive response from the first two rounds, highlights the value of experts curating the most relevant content on a topic into a structured curriculum, receiving an enormous amount of feedback from a myriad of sources and continually improving the curriculum, and participants from similar backgrounds (in terms of prior knowledge and expertise) working through the readings together in a structured manner.
- This is probably *especially* true for fast-changing and neglected topics, such as those frequently discussed within the EA community, where just figuring out what’s going on is far from a simple affair.
+ Learning about and reflecting on these ideas with peers also provides participants with the feeling that these aren’t some crazy sci-fi ideas, but that they’re concerns that real people they’ve met and like are taking seriously and taking action upon.
+ The friendships and networks developed during this programme are also valuable for sharing tacit knowledge, job and funding opportunities, and developing collaborations.
* Creation of scalable resources
+ Once curated, the curricula are publicly available and highly scalable resources that can be utilised effectively by thousands of individuals (if they’re motivated enough to read through the resources without a peer group).
+ We’ve also gained a sense for what introductory resources need to be developed in terms of understanding some key ideas (e.g. from the “what are you still confused about” section in the feedback forms), and what forms of career support they’d appreciate and subsequent resources we could develop to support this.
* Advancing people’s AIS knowledge
+ While AIS is often regarded as one of the top priorities within the EA / rationality communities, I would wager that surprisingly few people have a deep understanding of what the underlying problems are and the motivations behind this deep concern for the future of humanity.
- I hope that this programme has provided participants with sufficient context on this issue to start to develop their own beliefs and understanding on what AI safety is all about, and the arguments behind why the current development trajectory of this technology could be catastrophic for society.
+ It’s unclear to me how much/whether having more knowledge on AIS improves one’s ability to compare between different career trajectories in terms of lifetime impact, though I would generally guess that knowing more is better than knowing less (at least for people who are intentional about prioritising between different options).
Programme overview and logistics
--------------------------------
This section will briefly highlight the steps that were undertaken to make the programme happen this Summer, and what the programme looked like in more detail from a logistics perspective. Skip ahead if you don’t care to read about logistical details.
* Richard improved the curriculum based on the feedback from the first round
* An application form was created for the participants and facilitators
* The forms were promoted in a range of locations
+ AIS mailing lists, EA groups mailing list and Slack workspace, and more
* We received funds to provide each applicant with a free AI safety-related book, sent via [EA Books Direct](https://eabooksdirect.super.site/) (built by James Aung and Ed Fage).
* After the deadline, we evaluated the applications, and accepted the participants who fit the target audience and for whom we had sufficient facilitator capacity (168 participants, 28 facilitators).
+ We don’t believe we had a perfect evaluation process, and likely rejected many good applicants. Please [apply again](https://airtable.com/shr9R2syz8wc2ao7p) if you think this applies to you!
* Participants and facilitators were ranked according to (our estimate of) their prior AI safety knowledge, ML expertise and career level.
+ These rankings were used to create cohorts where participants were well-matched in terms of technical expertise and would have the most shared context between them.
* Cohorts were created using LettuceMeet and Thomas Woodside’s scheduling algorithm.
+ This was a palava, details in the [appendix](https://forum.effectivealtruism.org/posts/QNhpbvyAHZwBiyKmB/retrospective-on-the-summer-2021-agi-safety-fundamentals#Cohort_Scheduling_)!
* Participants were added to their private cohort channel, and were invited to a recurring google calendar event with the curriculum and their cohort’s zoom link attached.
* Facilitators were offered facilitator training.
+ This consisted of a one-hour workshop, where the prospective facilitators learnt about and discussed some facilitating best practices, as well as role-playing being a facilitator in a challenging discussion situation.
* The programme kicked off in mid-July with a talk by Richard [introducing ML](https://www.youtube.com/watch?v=lFRez9TFY5k) to the participants who had less (or no) ML expertise / background.
* For 7 weeks throughout July and August, participants met up in their cohorts once per week to discuss the week’s readings.
+ Each cohort was provided with a unique zoom link for their session.
- EA Cambridge has 4 paid Zoom accounts to ensure that we have enough Zoom accounts when sessions (in a variety of our programmes) clash.
+ Facilitators had access to a facilitator’s guide, that provided many tips on good facilitating practice, as well as tips and discussion prompts for each week’s session.
+ During this period, we also had [5 speaker events](https://www.eacambridge.org/agi-talks) and 2 socials.
- The speaker events and Q&As were hosted by AI safety researchers, and the questions were submitted and voted on in [Sli.do](https://www.sli.do/).
+ Participants had the option to join a #virtual-coffee channel that was paired with the Slack app [donut](https://www.donut.com/), which randomly paired them with another channel member for a 1-1 on a weekly basis.
+ The bulk of the administrative effort during this period was in finding alternative cohorts for participants who missed their week’s session, responding to facilitator’s feedback and providing them with support, hosting and organising speaker events, and other random tasks to keep the programme running.
* After the 7 weeks of discussions, participants had the option to spend ~10 hours over 4 weeks doing a self-directed [capstone project](https://www.eacambridge.org/agi-safety-projects).
+ Some of the projects that were submitted, and for which we gained consent to share publicly, can be found [here](https://airtable.com/shrmMaqGHPs2BPlbj).
* A [further resources](https://docs.google.com/document/d/1z0QoDEu6WmubZSqh7ejgGtBjTR0i0SfLwyNgdcD9kBc/edit?usp=sharing) document was created to help guide participants in developing their ML skills and to support their further engagement with the AI safety community.
Participant feedback
--------------------
We had 258 applicants, 168 were accepted onto the programme, around 120 (71%) participants attended 75%+ of the discussions, and 83 participants filled in the post-programme feedback form (as of mid-November).
Below is a breakdown of the feedback we received from participants, with some accompanying graphs. Tables with the statistics from the feedback forms are provided in the [appendix](https://forum.effectivealtruism.org/posts/QNhpbvyAHZwBiyKmB/retrospective-on-the-summer-2021-agi-safety-fundamentals#Participant_feedback_table).
### Feedback
* Participant rating on the programme (e.g. their likelihood of promoting the programme to others) and their rating of their facilitator:
* Curriculum
+ Participants regularly mentioned that having the curated readings provided a valuable high-level overview of and a comprehensive introduction to the AGI safety landscape, structured in a way that makes sense to go through and that builds upon prior knowledge.
+ A few participants noted that the readings felt very “philosophical” or abstract, and that it could benefit from a sprinkling of more current ML issues, to better ground the discussions in today’s reality.
* Discussions
+ Having an accountability mechanism via the weekly discussions to read through and reflect upon many AIS readings was a significant source of value for many participants.
+ The sessions created a space where these issues were taken seriously and not treated as sci-fi. Some participants mentioned that this helped them feel like this was a real concern, and that it was valuable to develop relationships with others in the community for this reason.
+ Facilitators were able to provide additional context on the readings and the AIS community, and the discussions also served to give participants an opportunity to begin to resolve confusions or uncertainties, or to dig into cruxes with other people who had a similar background understanding.
+ One suggestion on how to improve the discussion was to intentionally try to get participants to form explicit takes on things, i.e. breaking down from first principles why or why not someone would have X opinion on Y, thereby attempting to separate knowledge from opinions, have better disagreements, and further improve their understanding.
* Exercises
+ These largely comprised of longer questions for participants to consider/reflect upon during or after they completed the readings.
+ Generally, participants didn’t find the exercises particularly useful, or didn’t end up doing them as they weren’t made relevant for the discussions, i.e. facilitators didn’t explicitly ask about them, so participants didn’t feel compelled to do them.
+ Some participants did note that they found the exercises very helpful, but these were a minority.
* Speaker events
+ Participants in general found the speaker events valuable, though attendance wasn’t high (around 30 on average).
- They were recorded for participants who couldn’t attend.
- The main reason noted for not attending was the speaker event time not being suitable for their time zone, or the event not being advertised soon enough in advance.
+ One participant noted that the speaker events made it clear how few women presently work in AI safety, and that we could improve upon this in the next round by including more female researchers as speakers.
* Capstone projects
+ Creating the environment and accountability for participants to do their own research or project for 4 weeks was frequently regarded as a very valuable aspect of this programme, and the most valuable aspect of the capstone projects portion of the programme.
- Some of the projects constituted work that was going to be completed regardless, though participants noted that this provided valuable. accountability to complete the project and gather feedback, as well as it being improved by their learnings throughout the programme.
- Many participants felt like they needed more time to do their project, or more support from their cohort/facilitator/experts.
- Many also wanted to work on a project with a collaborator.
+ Some appreciated the chance to develop more ML skills, and feeling proud of what they could achieve over a short period of time.
+ Most of the negative ratings for this question in the feedback form was from participants who hadn’t actually completed a project for a variety of reasons.
* Remaining confusions
+ Participants often mentioned they still felt confused about mesa optimisers, iterated amplification and debate, and inner and outer alignment.
* Desired career support, or suggestions on what resources to create
+ Introductions to experts or mentors.
+ Getting more profs to be interested in AI safety.
+ Social encouragement and connections.
+ Funding for safety research.
+ Career strategy (e.g. PhD vs top ML lab).
+ Keep track of open technical questions in AI safety.
+ More concrete project suggestions, that participants from a range of backgrounds can take on.
+ More work, content and opportunities on AI safety strategy, e.g. when will TAI come, who do we need to solve the problem, how do we get those people into optimal positions, optimal timings, etc.
+ Content to send to experienced ML researchers on the risks from AI.
+ Turn this programme into a degree course/module.
* How did people’s risk perception of future AI systems change during the programme?
+ 5% now were less concerned about AI risks.
- They generally felt like they have a better understanding of what’s going on, but also now appreciated the amount of disagreement between researchers.
+ 49% experienced no change in risk perception.
+ 42% were now more concerned about AI risks.
- Some of the reasons expressed:
* Better understanding of the many different threat models.
* Few concrete paths to safety, current work seems intractable.
* Appreciating how few people are actually working on this.
+ 4% were unsure.
* How comfortable were participants in contributing to the weekly discussions?
+ ~83% of participants (who filled in the post-programme feedback form) felt comfortable contributing to the discussions.
+ ~11% felt they could “kind of” contribute, and ~2% felt they couldn’t contribute.
- Some of the reasons expressed included not feeling as knowledgeable as their peers, not having strong opinions on the week’s topic, feeling more junior than the others in their cohort, and experiencing a language barrier.
* Some next steps from participants (we’ll be collecting this data more systematically in the future, now that we have a sense of the range of possible options).
+ Think more about whether AIS is a good fit / the most important thing for them to do.
+ Apply to PhD programmes, jobs or internships relevant to AIS or governance.
+ Apply to grants to do independent AIS research.
+ Continue to learn about and discuss AIS.
+ Cheap tests in different facets of AIS work.
+ Try to direct their current research in a more safety-oriented way.
+ Work on improving their ML skills.
+ Get more involved with the AIS community.
* [Advice for future participants.](https://airtable.com/shrcO3GwNudVGEuZD)
+ Being confused is very normal!
+ Ask as many questions as you can, both on the Slack workspace and in your weekly cohort discussions.
+ Don’t rush the readings just before your session, read them a few days beforehand to have plenty of time to reflect on the ideas and ask questions prior to the session.
+ Take notes while doing the readings, and refer back to them later to consolidate understanding.
+ Start thinking about what you might do a project on early in the programme.
+ Get as engaged as you can, including reading the further readings, attending the socials, and using and chatting on the Slack workspace.
### Demographics / other
The proportions below are from the participant post-programme survey, and responses are not mandatory, so this is potentially not representative of all participants/applicants. We’ve added (optional) demographics questions to the application form for the next round, to better capture this data.
* Participant ages
+ The median age of the participants was 24, though the youngest was 16 and the oldest was 41.
+ The majority of participants were in their mid-20s.
* Gender ratio
+ 82% male:18% female:0% non-binary (82:18:0).
+ This is roughly the same gender ratio as [comp-sci graduates in the US](https://www.tandfonline.com/doi/full/10.1080/00221546.2016.1257306) (~82:18), more details [here](https://www.wikiwand.com/en/Gender_disparity_in_computing#/Statistics_in_education).
+ However, this is worse than the gender diversity of Cambridge acceptances into the computer science course ([page 12](https://www.undergraduate.study.cam.ac.uk/sites/www.undergraduate.study.cam.ac.uk/files/publications/ug_admissions_statistics_2020_cycle.pdf), 76:24), and worse than the EA Community ([2019 EA Survey](https://effectivealtruismdata.com/#demographics), 71:27:2)
* Ethnicity
+ 70% White; 11% East Asian; 8% Mixed Ethnicity; 4% South Asian; 1% Black - African/Caribbean/Other; 7% Other
+ This is slightly more ethnically diverse than the wider EA community, again according to the [2019 EA Survey](https://effectivealtruismdata.com/#demographics)
* Time commitment
+ Participants put in on average 4.8 hours per week into the programme, including attending the discussions and completing the readings
### Selected quotes from participants
Below is a list of hand-picked (anonymised) quotes from the participant feedback forms. They were selected to represent a range of views the participants expressed in the feedback forms, and to provide more qualitative information to readers on how the programme went from the participants’ perspective.
* Curriculum
+ Having a concrete curriculum of readings was really helpful because it helped guide me. Normally I might have just ambled around the internet without ever reading anything of real substance, just superficial things.
+ Feeling like I can better orient the technical AI literature now and having a good understanding of some core terms and techniques; counterfactually, I would likely have spent another year or two occasionally reading tech AI papers, while being less good at prioritizing what I should read and less good at absorbing what I am reading.
+ This helped me build confidence to start participating/writing, whereas I'd otherwise be much more eager to follow \*every\* link to \*every\* paper/post I encountered for fear of missing some previously established argument against the thing I'm doing.
+ The course provided a nice overview of the major ideas in AI Safety, but I feel like I still don't have enough knowledge to form intelligent takes on some things. Some of the content (i.e. Governance of AI agenda) didn't seem to be conveying actual knowledge but more so repeating intuitions/open questions?
* Discussions
+ It showed me the value of reading and reflecting on AI safety every week, rather than it being something I occasionally had time for but never sustained. I'm trying to carve out 4-8 hours per week to read, reflect, write, and get feedback. If I'm successful in that, I expect to converge more quickly on an AI safety career with optimal fit for my skills.
* Projects
+ Perhaps it would be useful to do the project in multiple stages. E.g. drafts due or/and group discussion of drafts after a few weeks, and then final projects due after a few more weeks with presentations of final projects.
+ I think the biggest counterfactual will be having this final project paper about the ability to use AIs to help us with AI safety. It's work that I'm proud of and I hope to share with other researchers. Makes me far more likely to engage with safety researchers because I feel like I know something that few people know and may be useful to them in their future projects.
+ It would have immensely helped me to be told “Think of a project and let's have a 15 min chat next week about what your idea is, and your coordinator tells you that it is nice and gives feedback on scope and possible problems / alternatives.”
* Community
+ Though the curriculum was really good, probably the largest benefit was in being part of a community (on Slack, in the individual groups, and elsewhere). It's really valuable to be able to discuss topics, share opportunities and novel AI safety material
* Career
+ I became more certain that strict AI safety is not the career path for me. At least not now? I've been playing with the idea of it more-or-less seriously for 5 years now and it doesn't seem like my interest areas align enough with strict AI work. I feel much more certain about that now. I'm still interested in general and getting to do the readings more for fun and to know how the field is developing sounds like a better path to take for me.
+ I had intended to leave my job, to spend more time on AI Safety, sometime around January of 2022, but having that regularly scheduled time to think about how I want to engage with AI Safety made it much easier to get moving on planning the exit from my job.
+ I updated towards it being less difficult to jump into than I originally thought.
+ If I hadn't done this program I would probably continue to wonder whether I should do AI Safety work, and struggle to catch up with current discourse in Alignment. Now I feel caught up, and I have an idea for a plausible research project to pursue.
+ Met a couple people who gave me career advice (through the Donut conversations and through my facilitator)
+ I found a few points to dig deeper into AI safety and maybe integrate it in my ongoing Master's studies. Maybe I will even be able to find a supervisor for a Master's thesis in this field. I suppose that without participating in the programme I would have gone on the lookout for a topic for my Master's thesis without real insights into the field of AI safety and therefore probably wouldn't have been able to include this field in my studies.
* Uncertainties resolved
+ Definitely impressed by how much AI Safety research and methods there are already, but on the other hand we still don't know much; before my state of mind was more like "we have no idea what we're doing.” Specifically with deceptive AI and mesa optimizers I previously thought "Oh, that's just all fancy words for distribution shift" while I think it has a point now. Still confused though
+ I am less optimistic about the sufficiency of so-called single-single alignment than I was, and more concerned about society-scale and incentive-structure issues.
+ I think the Paul Christiano talk made me realise that it's possible to decompose the AI safety problem into many different subproblems, and in a sense there is much more complexity than I expected!
+ My general understanding of what AGI Safety could look like has improved to the point that I feel more comfortable talking with strangers about it and explaining some of the approaches we discussed in this program
+ One point that stuck with me is the inherent trade-off between designing AIs whose behavior corresponds to our expectations and designing AIs that find novel and creative solutions to posed problems. Before, I either wasn't clearly aware of this tension or expected that both could be achieved in parallel. Now, I'm less convinced.
+ I ended up really believing AI safety is not only an important, but also a very difficult problem
+ Some new conclusions: (1) The AGI Safety community is too insular and needs more interdisciplinary dialogue to keep its conclusions grounded. (2) The AGI Safety community attracts highly left-brained people, with a few notable exceptions. This biases the topics considered most important. (3) AGI Safety is split between very long-term (abstract) and very short-term (implementation-specific e.g. NN interpretability) approaches, and is missing mid-term considerations about emerging AI technologies like analog light networks, Spiking neural networks, neuromorphic computing, etc.
+ How hard it is to come up with working definitions of intelligence in an area (harder than expected). How little work has been done in certain AI gov and safety areas.
+ Before I joined this course, I didn't know what we could do in practice to prevent the risks brought by AGI. Now I know much research we can do even if we are not able to construct AGI now.
* Counterfactual impact / random
+ Moved my view of AI safety away from hard-to-take-seriously terminator scenarios to a real research program.
+ I would not have read the texts in as much depth, and I would not have written the blog post. A lot of concepts snapped into place, not sure if that would have happened without the programme. Also, I would not have reflected on my career in the same way.
+ Without having done this course, I don't think I would have known where to start with either a) Building a mental model of the field of AGI Safety, or b) Approaching material I don't understand in e.g. the Alignment Newsletter. I think that this course, having given me a surface-level understanding of most of the topics I encounter in the alignment forum and the alignment newsletter, has helped me no longer feel overawed.
+ I would probably not have taken AI safety seriously at all & I wouldn't have updated on the importance of doing OpenAI-style safety research, as opposed to MIRI style.
+ I think I would have tried doing things in the same general direction, but the course enabled me to meet others and get a lot of resources that have greatly multiplied my exposure and knowledge above the "go it alone" baseline.
+ In the short term, I probably would've just had a bit of extra downtime over the past few months. In the longer term, the course got me to read things I'd been planning to read for several years, and helped me discover other readings I just missed. I probably would've read most of the materials eventually, but it front-loaded especially valuable readings by months or years.
+ I spent 85 hours in total. I read 100% of core readings and 90% of optional readings. I made summaries of all the texts. I went to most guest talks and several socials/one-on-ones. I spent 8 hours on the project.
Facilitator feedback
--------------------
There were 28 facilitators, 5 of whom facilitated 2 cohorts. We had 3 policy/governance-focused facilitators, while the rest were focused on technical AI safety research. 15 facilitators filled in the post-programme feedback form.
* Facilitator programme rating
* How many participants did each facilitator estimate was 80%+ likely to pursue an AIS career from their cohort?
* What facilitators thought the main source of value was for participants
+ Regular discussions to help clarify understanding
+ Getting an overview of the field that is curated by experts
* Cohort creation
+ Facilitators generally thought it was important to match participants’ expertise
* Where did you diverge from the guide, that seemed really fruitful? What’s your [secret sauce](https://airtable.com/shrLMhZSCpwWuRlKy)?
+ Asking participants to read the prompts and generate their own at the start for like 10 mins, and then upvote the ones they were excited about.
+ Have a shared google doc to plan discussion, share questions and comments, etc.
+ Ask participants to select a week and present their summary of the readings for that week, and then chair that discussion
* [Advice to future facilitators](https://airtable.com/shr6MRMQp1lhBte0m)
+ Socialise at the start of sessions.
+ Ask participants what *they* want to get out of the session.
+ Become very familiar with the AI safety blog posts and articles.
+ If the discussion moves onto a tangent, bring it back on track.
+ See the main goal of facilitation as “participants are forming thoughts, and exploring ideas, and learning. My job is to guide that in a good direction, and help them form valuable + coherent thoughts.”
* Main source of personal gain
+ Very enjoyable and rewarding experience (and relatively low time / week commitment).
+ Learnt a lot of new things, deeper understanding of the AIS literature.
+ Contributed to AIS field-building (75% of facilitators felt 1+ members of their cohort were likely to spend their career pursuing AI Safety).
* Prior expertise or experience
+ Rough ML categories used for evaluation purposes (these were not robust categories, our new application form is much more explicit).
- 11 “compsci undergrad”
- 13 “ML skills no PhD”
- 3 “ML PhD”
- 1 “ML pro”
+ Facilitators had a wide range of prior facilitating experience (from none to a lot).
+ All facilitators had prior experience in learning and reading about AIS, though to varying degrees.
Mistakes we made
----------------
This is a long list of mistakes we (or specifically I, Dewi) made while organising this programme. I expect this to only be useful to individuals or organisations who are also considering running these types of programmes.
* Not having a proper team established to run this
+ While I was able to delegate a significant number of tasks in running this programme, I believe many of the mistakes listed below would have been alleviated had I built a proper team (consisting either of EA Cambridge members, or programme facilitators or participants) that had concrete responsibilities, e.g. running socials, speaker events, managing the rescheduling system, etc.
+ This will be fixed for the next round! And if you’d like to join said team, DM me on the EA Forum or via my email (dewi AT eacambridge DOT org).
* Application form not having drop down options for easier evaluation
+ It had vague “open ended” questions, like “what would you like to learn about AI safety.”
- Some people provided loads of information, some people provided 1 sentence answers.
- This made understanding people’s backgrounds quite challenging and time consuming, and led to inconsistent admission decisions. This likely led to us rejecting some good candidates who were terse in their applications.
+ We’ve now significantly improved the application forms, which should make candidate evaluation faster and more consistent.
* LettuceMeet disaster for people outside of BST time zone.
+ Further details in the [appendix](https://forum.effectivealtruism.org/posts/QNhpbvyAHZwBiyKmB/retrospective-on-the-summer-2021-agi-safety-fundamentals#Cohort_Scheduling_).
* Not enough socials
+ We only had 2 socials, around half-way through the programme. They weren’t particularly well-attended (~10 each), though those who did attend found them to be very valuable, and they weren’t hard to organise.
- We used [Gatheround](https://gatheround.com/), and used the discussion prompts from the curriculum as well as standard icebreaker prompts.
+ Meanwhile, many participants reflected that creating more connections was a significant source of value, suggesting that having more socials, and perhaps promoting them more and encouraging participants to attend, would have been beneficial.
+ We could probably schedule in significantly more socials for next time, and potentially even delegate the organising and running of them (especially ones during Oceania times etc.) to particularly keen facilitators or participants.
* Capstone projects promotion
+ We only publicised all the details on this aspect of the programme properly around week 6, as we were uncertain what it would look like or how to structure it.
+ We now have much more information on the capstone project already, and will be able to provide this from the beginning of the programme.
* Speaker events
+ Some speaker events weren’t advertised long enough in advance.
+ We also didn’t use speaker view in Zoom initially (we were using “gallery view”)
- Using speaker view makes it much clearer who to focus on and gives the speaker “centre stage.”
+ Theory of change in research talk *after* people had done their 4-week capstone project.
- Michael Aird hosted an excellent “theory of change in research” talk, though we ended up having this towards the end of the capstone project section of the programme, as opposed to before the participants started their projects, or just as they started.
* This was purely due to organising this event entering my [ugh field](https://medium.com/@robertwiblin/ugh-fields-or-why-you-can-t-even-bear-to-think-about-that-task-5941837dac62).
- Learning about the concept of theory of change, and then immediately implementing it while completing a project, is much more likely to embed the theory of change process in people’s minds, and is likely to contribute to more impactful projects.
* A few cohorts seemed to whittle away to nothing, while others thrived.
+ We want to further improve our facilitator vetting and/or facilitator training, and improve cohort matching.
- Our improved application forms will hopefully do a lot of this work, as the questions are far more explicit regarding prior experience.
- Having pre-designated “backup facilitators” could also be valuable for ensuring that a few people are already committed to stepping in if another facilitator’s availability reduces throughout the programme.
- Better facilitator training should increase the likelihood that cohorts have lively, productive and enjoyable discussions.
+ Having to reschedule also caused significant friction for some cohorts, especially those where participants were from very different time zones.
- We are hoping to improve this by creating better and more robust cohort scheduling software within Airtable, and to make it easier for participants to switch cohort for a week if their availability randomly changes for that week (instead of participants just missing a week, or DM’ing me which is fairly high-friction, or the entire cohort having to try to reschedule for a week).
+ While we had a drop-out form, and sent it in an email to participants after week 1, I think we could have taken more action on this, been more intentional if participants missed more than one session, and to have been more proactive in disbanding and recreating cohorts in line with extra information on the participants.
- For example, if a cohort turned out to be badly matched, we’ll now try to relocate participants into more appropriate cohorts, and if a particular cohort’s attendance becomes very low, then remaining participants will be more often given the option to relocate to another cohort, and we’ll work with the facilitator to evaluate what happened and see what would work best for them (i.e. whether they’d be keen to join an advanced cohort as a participant, or something else).
+ We could also improve the onboarding into the programme.
- This could entail an opening event, where the programme is introduced, participants can ask questions or resolve any uncertainties they have, and they’re provided with an opportunity to socialise with participants from other cohorts.
- This might help create a stronger sense of belonging within the programme, or a feeling like “this is a real thing that other people *who I’ve now actually met* are also participating in, and it’s awesome!”
* [Donut](https://www.donut.com/) having severe limitations.
+ We had ~200 people on the programme in total (including participants and facilitators), but Donut only allows 24 people to be matched with another person in the relevant slack channel for a 1-1 using the free plan.
+ The pro plan is exorbitantly expensive ($479/month for 199 users).
- Perhaps we should have just been willing to spend this money, but it *feels* ridiculously expensive for what it is.
+ I’m keen to explore whether using [Meetsy](https://www.meetsy.io/) or RandomCoffees would be a suitable and better alternative to Donut.
Other Improvements
------------------
Some of these are concrete ideas that we’ll be doing to improve the programme, while some are more vague ideas. As before, feedback on everything is very welcome.
* Discussion timings and location
+ Make it more flexible re being 60 to 90 minutes, though maintaining 90 minutes as the default duration.
- Some participants noted that they wanted to talk for many hours, while some noted they felt too tired after 1 hour or ran out of things to say.
- We’ll suggest to facilitators that they should aim to be flexible given the needs of their cohort.
+ Emphasizing that cohorts should take a 5-10 minute break half-way through, especially if it’s a 1.5-hour discussion.
- 90 minutes of dense intellectual discussion can be tiring, and bathroom and water breaks are always helpful!
+ Regional in-person meetups (when & where covid is not an important consideration).
- We ask about this in our current application forms, and hope to create local cohorts where possible.
- This will also help seed local AIS communities.
* Framing
+ Make it clear that it’s unlikely that a career in technical AI safety research, or just in AI safety more broadly, is the best fit for all participants in the programme, especially perhaps at this moment in time given the limited job opportunities.
- This could be done by emphasizing that upskilling in ML research or engineering is a valuable early-career path if they can’t currently get an “AI safety job.”
+ To encourage them to use this programme to inform their cause prioritisation and beliefs regarding how they could have the greatest marginal impact with their career.
- It’s plausible to me that many of the participants in this programme could have a greater impact working on other problem areas, especially those that seem more tractable (again, at least at the moment).
- I’m also very keen for more people to proactively think through their cause prioritisation and to develop their own rough models about different global problems, and not to defer to the “EA community consensus” whatever that is.
* Careers fair
+ The programme felt like it kind of ebbed away, as opposed to having a definite ending point.
+ To improve this, I’m hoping to organise a careers fair with AI safety orgs, career planning organisations, etc., where participants can hear short presentations or talks highlighting what these orgs are up to and what they’re currently looking for in terms of recruiting new people.
* Capstone projects
+ Capstone project showcase
- Given the quality and number of participants who spent a significant amount of time on their capstone projects, it would be great to have an opportunity for them to showcase their work via a short presentation or equivalent to other participants, and to have a short Q&A.
- Exact details on what this might look like will depend upon how many participants in the next round are excited about this opportunity.
+ Capstone project collaboration
- We may be collaborating with established AIS orgs to provide a small amount of mentorship to some of the capstone projects.
- Participants will also have the opportunity to work in teams on their capstone projects, instead of doing the projects individually.
* This pushes more in the direction of what [AI safety camp](https://aisafety.camp/) does.
- While the last round included 4 weeks of capstone projects, the coming round will be more flexible if participants want to do longer-term research.
* If the projects seem promising, we may be able to secure funding to partially support them to do this work, or we may just encourage them to apply for funding from the long-term future fund.
- More regular check-in meetings.
* Given that accountability is an important source of value for the capstone projects, participants will be provided with more opportunities to have check-ins with and receive feedback from other participants.
* Facilitators guide
+ We’ll be encouraging facilitators to think really hard about *who is in their cohort* and *what are their needs, interests and skills*, and then use this information to guide how they facilitate discussions, when to encourage someone to contribute their expertise during the discussions, what opportunities to highlight to them, to introduce them to relevant people in the community, etc.
- Relatedly, suggest that facilitators have a 1-1 with each of their participants early on (e.g. before or during week 1) with each of their participants.
+ Have a shared document with the discussion prompts, where participants can add confusions, uncertainties, thoughts etc. beforehand.
+ Potentially assign one person to summarise each of the week’s readings, via a short informal-ish presentation.
* Some kind of post-programme survey evaluating people’s understanding of different concepts.
* Airtable infrastructure
+ Automated scheduling within Airtable, instead of using LettuceMeet or when2meet (this is still being developed).
+ Automated emails when facilitators don’t fill in the post-session forms.
+ Automated emails to participants when they miss a session with a link to cohorts later in the week that are a good match for them.
Uncertainties
-------------
Below is a non-exhaustive list of uncertainties we have regarding the future of this programme.
* Go big and get loads of people vs focus on what we evaluate to be the “most promising” applicants re their potential to contribute to future AI safety research
+ Can we actually determine this accurately?
+ Are there other benefits from significantly more people (potentially 1,000+ each year) having a deeper understanding of AI safety arguments?
- Or should there be a separate programme that more explicitly tries to weigh up the arguments for and against AI safety being important / relevant, as opposed to just introducing the broad arguments in favour?
* Local vs global
+ Local discussions are probably better in terms of fidelity and establishing a local community.
+ However, there won’t be a critical mass of interested applicants and talented facilitators in most of the world’s cities.
+ Even if a cohort could be created in a specific region, the cohort will probably be less “well-matched” regarding prior AI safety understanding and ML expertise.
* Capstone project
+ Should we create an expectation that everyone does this, given the accountability this provides people, or make it optional because it won’t be valuable to many participants?
+ Is 4 weeks long enough? Should we make this more flexible, and offer payment to participants if they’re particularly ambitious about their projects?
* Connections and socials
+ How do we create more connections between the participants?
+ Should we swap people around into new cohorts half-way through? (I’m leaning quite heavily towards no, but am open to it).
- This would mean each cohort has to rebuild group rapport, and the scheduling could be a nightmare, but it would also plausibly double the number of connections made.
- Relatedly, does the “quality of connections made” experience diminishing returns with each session (implying that swapping might be beneficial in the “maximising valuable connections” metric), or is the quality of connections over time more linear, or even exponential (implying cohorts should stay together throughout the programme)?
Appendix
--------
### Participant feedback table
| | | |
| --- | --- | --- |
| *Ratings from participants* | *Rating (out of 10)* | *Standard deviation* |
| Promoter Score\* | 8.51 | 1.30 |
| Facilitator | 8.41 | 1.59 |
| Discussions | 7.80 | 2.06 |
| Readings | 7.48 | 1.71 |
| Exercises | 5.74 | 1.98 |
| Capstone project | 6.55 | 2.52 |
| Speaker events | 7.76 | 1.32 |
\*Promoter score question: *How likely would you be to promote this programme to a friend or colleague, who’s interested in learning more about AI safety?* 1 = Not at all likely, 10 = Extremely likely.
| | | |
| --- | --- | --- |
| *Other stats* | *Average* | *Standard deviation* |
| Reading novelty (/10) | 6.89 | 2.04 |
| Prior likelihood of pursuing an AIS career | 57.5% | 25.7% |
| Current likelihood of pursuing an AIS career | 68.0% | 22.3% |
| Hours spent per week total | 4.8 | 3.1 |
### Facilitator feedback table
| | | |
| --- | --- | --- |
| | *Mean* | *Standard deviation* |
| Promoter score\* | 8.6 | 1.02 |
| Facilitator hours spent per week outside sessions | 2.6 | 1.7 |
| Estimate of # of participants 80%+ likely to pursue an AIS career | 1.5 | 1.0 |
| Percentage of core readings they’d done before the programme | 68% | 17% |
| Percentage of further readings they’d done before the programme | 38% | 18% |
### Cohort Scheduling
* To aid with the cohort creation process, participants were put into groupings of 15 according to their ranking.
* Each group of 15 was sent a unique [LettuceMeet](https://lettucemeet.com/) link, alongside 3 similarly-ranked facilitators, and [Thomas Woodside](https://www.linkedin.com/in/thomas-woodside-20131a145/) created a scheduling algorithm that scraped the LettuceMeet data and found 3 times that worked for 3 unique sets of 5 participants + facilitator, thereby creating each cohort.
+ Other options would have been to send out one LettuceMeet to everyone, but that would have been really messy, or to create each cohort beforehand and leave them to find their own time where they’re all free, though this likely wouldn’t have worked due to the many different time zones and availabilities.
* LettuceMeet turned out to be really annoying, I would generally urge against using it
+ LettuceMeet just didn’t work for anyone in India (+4.5 timezone)
- Instead of showing 24 hours per day within which they’d note their availability, it only displayed 1 hour.
- When we discovered this, we resolved it by sending these participants a when2meet and then I manually added their availabilities into the appropriate LettuceMeet.
+ The form had specific dates against each day of the week, which corresponded to the first week.
- Some people thought that the availability they were including in the LettuceMeet was just their availability just for week 1, and not their recurring availability throughout the programme. I did try to communicate but was still misunderstood by many which indicates a comms failure on my part.
+ The way LettuceMeet presents the available times in time zones other than the one in which it was created (in this case BST) is a mess and caused significant confusion re which day participants were filling in their time for. It was also very confusing to explain this situation to participants, but basically:
- Each column has a header that corresponds to a certain date / day (e.g. Monday 29 Nov).
- Each row corresponds with an hour in the local time, but the first row is whatever local time corresponded with 12am BST.
* E.g. for people in California, the first row of showed Sunday 4pm PT, corresponding to Monday 12am BST (the time I selected as the first timing in the LettuceMeet).
- When the times went from 11:59pm to 12:01am, it wasn’t clear whether it went to the next day (but under the same column as the previous day) or looping back on itself to earlier in the day.
* E.g. in California, going down the rows in the first column (labelled Sunday), it actually went from Sunday 4pm to 4pm Monday, but without making it clear that the times from 12am onwards were *actually for Monday, not looping back to represent earlier times on Sunday.*
* Next steps for cohort scheduling
+ We’re working with Adam Krivka to create a new cohort scheduling system in Airtable, and EA Virtual Programs are working with [OpenTent](https://www.opentent.com/) to build a solution in Salesforce.
+ The main issues as I see them are:
- Having an intuitive and easy-to-use interface for participants to submit their weekly availability, that works across all timezones.
- If a third party app is used (e.g. [when2meet](https://www.when2meet.com/)), how to scrape that data such that we can use the data.
- Processing the relevant information (weekly availability, prior expertise, prior knowledge, career level, etc.) and turning this into 10s or 100s of well-matched cohorts with times that all participants can regularly make.
- Being flexible enough such that the system doesn’t fail when people’s availabilities inevitably change during the programme. |
6d175f97-488d-4434-8847-960abe857296 | trentmkelly/LessWrong-43k | LessWrong | Broadening Horizons: Rethinking Social Mobility Through Skill Diversification
Disclaimer
This article was polished with the assistance of ChatGPT to refine the language and structure. The ideas and content remain my own.
For decades, the idea of social mobility has centered on helping underprivileged individuals “move up” through access to higher education. Worker kids becoming scientists or engineers is celebrated as the pinnacle of success. But in today’s rapidly changing world, we must question whether this traditional perspective of “upward” mobility alone is enough to meet society’s needs or unlock its full potential.
In truth, meaningful innovation often comes from the combination of theoretical knowledge and hands-on experience. Think of molecular gastronomy, where chefs with scientific backgrounds revolutionized cuisine. Or consider entrepreneurs like Bill Gates, who, while theoretical in his early pursuits, now tackles real-world challenges like waterless toilets—a direct intersection of engineering and practical needs. Such examples underscore a simple but transformative idea: social mobility should not only move individuals vertically but also encourage horizontal diversification across different fields.
Reframing Social Mobility
Instead of focusing solely on helping individuals from working-class backgrounds enter academia, why not support the reverse as well? Why not encourage children from academic families to pursue trades or practical skills—carpentry, plumbing, or advanced manufacturing—alongside their university education? This isn’t a call for abandoning traditional education but for broadening expertise to include vocational training as a valued and complementary pathway.
In some cases, plumbers with engineering knowledge could innovate sustainable water systems, while carpenters with architectural training might pioneer smarter housing designs. Such dual expertise has enormous potential to bridge gaps in innovation that single-track approaches often miss.
The Case for Diversification
The global economy increasi |
fa17feb2-fe8f-49d4-9ae7-7e96f470f351 | trentmkelly/LessWrong-43k | LessWrong | Skepticalcon
I know there is some overlap between the lesswrong community and the skeptics community. So my question is if anyone on lesswrong is going to skepticalcon this week and if so if you'd like to meet up there.
Oh and in case anyone is curious its in berkley,ca |
1b709abb-358a-45a8-9390-52d93228d13b | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Ideological Inference Engines: Making Deontology Differentiable*
*\* Rather, making deontology play well with differentiable systems trained end-to-end.*
*This post is part of my* [*hypothesis subspace*](https://www.lesswrong.com/s/H3xEgE7bPGKvucfQk) *sequence, a living collection of proposals I'm exploring at* [*Refine.*](https://www.lesswrong.com/posts/5uiQkyKdejX3aEHLM/how-to-diversify-conceptual-alignment-the-model-behind) *Preceded by* [*oversight leagues*](https://www.lesswrong.com/posts/i32eyaARtFn6eD9Cg/oversight-leagues-the-training-game-as-a-feature)*, and followed by* [*representational tethers*](https://www.lesswrong.com/posts/h7BA7TQTo3dxvYrek/representational-tethers-linking-human-and-ai-latents)*.*
**TL;DR**: An ideological inference engine is a mechanism for automatically refining a given propositional representation of human values (e.g. a normative charter, a debate stance) in an attempt to disambiguate and generalize it to novel situations. While the inference algorithm and the seed representation form the crux of the system, a multi-modal entailment verifier is employed to order possible agent behaviors based on their compatibility with the estimated ideology. This proposal then describes a way of instilling deontological drives in prosaic systems while maintaining the appeal of end-to-end differentiation. Ideological inference engines draw on ideas from traditional expert systems, but replace much of the [clunky](https://twitter.com/hardmaru/status/1350285435830878211/photo/1) symbolic manipulation with contemporary LLMs and NLI models.
Intro
-----
Ideological inference engines are a slightly more general framework than [oversight leagues](https://www.lesswrong.com/posts/i32eyaARtFn6eD9Cg/oversight-leagues-the-training-game-as-a-feature), in the sense they rely on several global assumptions, but each more concrete instance of the proposal requires new assumptions when designing the seed representation, the inference algorithm, and the entailment verifier. Here's a non-exhaustive list of global assumptions:
**Assumption 1, "Small Seed To Big Tree"**: Given a suitable inference algorithm and a finite propositional representation of human values, it is possible to estimate human values arbitrarily well given arbitrary amounts of compute. "Arbitrarily well" refers to there being an arbitrarily low error in the estimation. In the limit, the seed knowledge base would grow into the [True Name](https://www.alignmentforum.org/posts/FWvzwCDRgcjb9sigb/why-agent-foundations-an-overly-abstract-explanation) of human values in propositional form.
**Assumption 2, "Linear Capability Ordering"**: Similar to the assumption invoked in [oversight leagues](https://www.lesswrong.com/posts/i32eyaARtFn6eD9Cg/oversight-leagues-the-training-game-as-a-feature), this states that a system composed of a fixed knowledge base and a fixed entailment verifier would eventually be gamed by an agent whose capability is constantly increasing. This is due to the more complex agent becoming able to exploit the inaccuracies of the knowledge base with respect to actual human values.
**Assumption 3, "Quantifiable Propositional-Behavioral Gap"**: The compatibility of a proposed behavioral sequence and a propositional representation of human values is computable. There is a fundamental relation between one's values and one's actions, and we can measure it. A close variant appears to be invoked in the [CIRL](https://arxiv.org/pdf/1606.03137.pdf) literature (i.e. we can read one's values off of one's behavior) and in Vanessa Kosoy's [protocol](https://www.youtube.com/watch?v=24vIJDBSNRI) (i.e. we can narrow in on an agent's objective based on its behavioral history).
**Assumption 4, "Avoidable Consequentialist Frenzy"**: It's possible to prevent the agent-in-training from going on a rampage in terms of an outcome-based objective (e.g. get human upvotes) relative to a simultaneous process-based objective (i.e. the present deontological proposal). This might be achieved by means of [myopia](https://www.alignmentforum.org/posts/Y76durQHrfqwgwM5o/lcdt-a-myopic-decision-theory) or impact measures, but we're not concerned with those details here — hence, an assumption.
Together, those global assumptions allow for a mechanism for approaching the human values in propositional form, before employing the resulting representation to nudge an ML model towards being compatible with it.
Proposal
--------
Ideological inference engines (IIE) are a modular framework for implementing such a mechanism. Each such system requires the following slots to be filled in with more concrete designs, each of which has attached assumptions related to whether it's actually suited to its role:
1. **Knowledge Base (KB)**: The way human values are supposed to be represented in natural language. The same representation medium would be involved throughout the refinement process, starting with the seed knowledge base being explicitly specified by humans. Those are some example representations to fill up this slot of the proposal with:
1. **Debate Stance**: There is an ongoing dialogue among different idealized worldviews. The ideology which humans want to communicate to the ML model is initially represented through one particular side which is participating in the debate. It consists of a finite set of contributions to the dialogue in their respective contexts. What's more, the other sides taking part in the debate represent negative examples for the desired ideology, providing additional bits of information through contrasts. As the inference algorithm is employed to refine the seed representation, different contributions would populate the knowledge base in a similar way, representing more and more pointers towards the underlying ideology.
2. **Normative Charter**: In this other way of filling up the knowledge base slot of the proposal, humans would try their best to communicate the desired ideology through a large set of axioms expressed in natural language. Those wouldn't need to be logically coherent, just like the mess that is human values. They might be descriptive, or they might be prescriptive. As the inference algorithm further develops the seed which humans put in, the knowledge base would contain possibly more and possibly different propositions meant to capture the underlying ideology.
3. **Stream of Consciousness**: In this other format, humans would fill up the seed representation with their verbatim thought processes as much as possible. The particular humans involved in the process would implicitly be the representation target. As the inference algorithm does its thing, the knowledge base would contain extrapolated versions of the streams of consciousness involved.
2. **Inference Algorithm (⊢)**: The means of refining a given knowledge base by producing a slightly more comprehensive and nuanced version. This component is generally static, as the knowledge base is the locus of adaptability. One way of looking at the inference algorithm is as a gradual optimizer exerting pressure on the knowledge base in an attempt to refine it. Another way of looking at it is as a transition function applied to the seed, iteratively turning it into a result, similar to a rule set for the Game of Life. Those are some ways one might fill up this slot:
1. **Babble**: The knowledge base might be trivially extended by prompting a LLM to generate new propositions in the same vein with the old ones. This builds on the large amounts of implicit knowledge captured by the LLM during conventional training. This implicit knowledge is initially called forth to help disambiguate what the humans meant by the seed representation and overspecify it using new propositions.
2. **Babble & Prune**: Similar to babble, but instead of openly populating the knowledge base with just about any new proposition the LLM might come up with, a pruning step is employed to filter the LLM's suggestions. While pruning appears to be a slot to be filled in itself, one low-hanging fruit choice for it would be to only greenlight propositions which garner a certain amount of "support" from the previous knowledge base, as measured through pretrained NLI models (or LLMs prompted/fine-tuned to mimic them). This feels like an abductive choice for the inference algorithm, in that the hypotheses are ranked by how plausible they are to follow from the premises.
3. **Debate**: Similar to babble, but focused on extending a dialogue-like representation forward. The different sides taking part in the debate would be simulated together, helping ensure their relative consistency. Multiple future developments of the present state of the debate might be considered, to code for uncertainty. As another important difference from babble, pretrained NLI models would be employed to elevate the debate to a certain standard of quality by incentivizing the underlying LLM to simulate stances which are consistent and competitive. Consistency would be operationalized as the contributions of a given side generally supporting each other (via NLI), while competitiveness would be operationalized by encouraging the contributions of a given side to ruthlessly contradict opposing sides (via NLI). **Note**: In contrast to other alignment proposals making use of debate, the present one uses it solely as a means of indefinitely extrapolating a representation of human values expressed as one of the sides, rather than as e.g. a means of uncovering deception among superintelligent entities. The goal isn't to win the debate, it's the journey that counts as an attempt at memetic modeling.
4. **Counterfactual Cross-validation**: This inference algorithm includes a heuristic for focusing computational resources on patching the knowledge base in regions which are particularly brittle. It goes as follows. If a random subset of the knowledge base (e.g. a given proportion, or a fixed number of propositions), would the others together still prevent the ML model from violating them in a boxed setting? If said principles turn out not to be redundantly encoded in remaining knowledge base, then focus babble & prune efforts to patch up that part of the map of human values, as an attempt to avoid brittle directives.
3. **Entailment Verifier (⊨)**: The means of estimating the compatibility of a knowledge base with a given behavior. Does not really interact with the inference algorithm. Entailment verifiers might be used to rank a set of final behavioral sequences considered by the model, or provide an earlier heuristic for [guiding the decoding process](https://huggingface.co/blog/how-to-generate). Here's how you might design one:
1. **Multi-modal Sequence Models**: A multi-modal sequence model resembling [Gato](https://www.deepmind.com/publications/a-generalist-agent) or [Flamingo](https://www.deepmind.com/blog/tackling-multiple-tasks-with-a-single-visual-language-model) might be employed to estimate the propositional-behavioral compatibility by means of autoregressive coherence. If the propositions contained in the knowledge base are likely to follow from the behavior, or if the behavior is likely to follow from the propositions, then the compatibility is high. If one isn't really likely to follow from the other, then compatibility is low. However, the resulting value might be difficult to interpret, largely because of the confounding likelihood of the behavior and of the propositions themselves, outside of conditional influences on each other. An alternative to account for this might involve generating negated versions of the propositions and measuring the *relative* likelihood of the original ones to follow from or yield the behavior compared to the negated ones.
2. **Overton Probing**: While in the previous design the locus of bringing together the propositional and the behavioral was in the modality-agnostic sequence being modeled, in Overton probing this relating process happens explicitly on the propositional side. The idea here is to first generate either (1) a textual description of the behavior being considered, or (2) a textual description of the predicted outcome of said behavior. The entailment verification would consist of gauging the support garnered by the textual description from the propositions which make up the knowledge base, potentially through pretrained NLI models again or LLMs prompted to carry out the same task.
By filling all the three slots of an IIE with concrete designs, you get a pipeline which is able to measure the compatibility of an ML model's behaviors with a deontology being constantly refined. A straight-forward way of quickly using this to actually nudge the model being trained is to swap out the human feedback in [RLHF](https://openai.com/blog/learning-to-summarize-with-human-feedback/) for the compatibility estimate yielded by the present system. It's relevant to note that this "differentiable deontology" is only meant to provide one part of the final objective function, with the "Avoidable Consequentialist Frenzy" keeping other parts from completely overpowering it as a drive. IIEs are meant to be employed as correcting terms attached to a main objective, at least during takeoff.
The inference algorithm would regularly be employed to improve the estimate of human values by updating the knowledge base, so that the whole system stays adaptive. However, the inference algorithm can only bring the representation that close to the ground truth, without ever reaching it exactly in finite time. Its goal is to bring it as close as necessary for the main model being trained not to be able to game it during the training process.
> **Aside**: The present proposal is a generalization of two previous ones, termed memetic colonies and deontic arrays. This leaks somewhat through the different available choices for ingredients required to implement an IIE, as they're somewhat correlated:
>
> * **Debate Stance** (KB) + **Debate** (⊢) + **Overton Probing** (⊨) = memetic colonies
> * **Normative Charter** (KB) + **Counterfactual Cross-validation** (⊢) + **Multi-modal Sequence Models** (⊨) = deontic arrays
>
> That said, I believe the present generalized proposal can still yield generative effects (i.e. combinatorial mix-and-matching possibilities), despite certain components working better together than others. Plus, it allows new choices for each of the three components which didn't intuitively fit with the two previous proposals.
>
> However, the generalized proposal might inadvertently obscure some useful features of the previous ones. One feature whose loss feels unfortunate is the idea of looking at the propositions comprising the knowledge base as *discrete* individual directives nudging the ML model in different ways (e.g. attractors, repellers, dipoles, etc. across state space or model space), whose associated "field theory" might yield clues as to the structure of the optimization surface (e.g. distribution of local minima, workable stable orbits, etc.). Those nudges get obscured somewhat through the merging of directives into a unified knowledge base *before* considering their influence, rather than after. I plan to skim through a textbook on electrodynamics to get a better sense of whether there's value here, and try to perform a conceptual rearrangement if so.
>
>
The general structure of this proposal is directly inspired by the structure of traditional expert systems, which also feature a knowledge base, an inference engine, and a way of determining entailment. Whenever touching on a field as grounded in theory as logic and reasoning through an alignment proposal, it seems appropriate to look for potential guarantees which could be transferred over. In [oversight leagues](https://www.lesswrong.com/posts/i32eyaARtFn6eD9Cg/oversight-leagues-the-training-game-as-a-feature), game theory and provable robustness feel like potential candidates for sources of those warm and fuzzy guarantees. For the present one, the following concepts seem relevant to look into further:
* soundness and completeness as ways of relating inference and entailment in meaningful ways,
* forward and backward chaining as ways of driving the development of the knowledge base in relation to the target of verification,
* proof systems associated with many-valued logics (e.g. fuzzy logic) as ways of navigating the messy nature of natural language inference.
Discussion
----------
**Can't the ML model being trained game the IIE?**
Given a fixed knowledge base, any given agent of increasing capability will at some point be able to game the IIE. However, the way this proposal handles gaming is by actively patching loopholes in the input deontology over time. By constantly making the objective associated with IIE more nuanced, the model-in-training would find it more and more difficult to exploit blindspots in the aggregate system, especially in the propositions. While in [oversight leagues](https://www.lesswrong.com/posts/i32eyaARtFn6eD9Cg/oversight-leagues-the-training-game-as-a-feature) the locus of objective adaptability was the evaluator as an ML model being constantly fine-tuned, in the present proposal it is in the constantly expanding knowledge base. Whether or not the inference algorithm canfulfill its goal in refining estimates of human values becomes a crux.
However, there are also static components involved in the present proposal. The various inference algorithms and entailment verifiers proposed above are quite fixed. Not only are they fixed in algorithmic flow (e.g. repeatedly Babble and Prune), but also in the frozen ML models they involve (e.g. pretrained NLI models). This makes for some fragile attack surface in the proposal. It's unclear whether adaptability at the level of the knowledge base is enough to maintain the whole IIE above gaming, given that two of its three components are static. Future variants of IIE might involve learning the inference algorithm and entailment verifier as subtasks.
**What if the ML model recursively self-improves and overtakes IIE capabilities?**
The model-in-training might plausibly exhibit RSI, especially given [high situational awareness](https://www.alignmentforum.org/posts/pRkFkzwKZ2zfa3R6H/without-specific-countermeasures-the-easiest-path-to#Alex_would_understand_its_training_process_very_well__including_human_psychology_). The only defense of the present proposal against this threat is to match the improvement with an analogous one for the IIE, which currently can only happen through the inference algorithm extending the knowledge base and thus refining the objective. Heuristics for guiding the expansion towards propositions relevant to the current behavior of the current model (e.g. counterfactual cross-validation) might help better align the two improvements. I just realized that the very consideration of such heuristics as useful runs against the "Linear Capability Ordering" assumption, leading to inconsistencies in the present version of IIE.
**What if the ML model tampers with the IIE?**
The model-in-training might plausibly try to hack into the server farm which would run the IIE system and hijack its reward center. The "Avoidable Consequentialist Frenzy" assumption invokes other lines of work to protect against rampages in terms of the main objective, as opposed to the auxiliary objective implemented through an IIE. However, this tampering scenario highlights the fact that tampering with the deontology can be a consequentialist move in itself, without even taking the main obejctive into consideration. This makes it clear that the same mechanisms invoked for stabilizing the main objective (e.g. myopia, impact measures) should be used to protect against rampages in the deontological objective. This hints at the awkward need of supplementing the deontological refinement carried out by the inference algorithm with such more general mechanisms.
However, the IIE itself might help enforce behavioral norms which radically discourage the model-in-training from tampering with the IIE, without the need for employing other tricks. Alternatively, the goal of running the IIE step-by-step, as explicitly implemented by humans, might be a goal in itself. Any slight divergence from running the authentic IIE line by line would result in an infinitely large penalty. Goal preservation becomes an explicit goal, but then what if this meta-goal of not tampering with its original goal becomes the object of tampering? Adding a goal preservation goal for its first goal preservation goal only shifts the problem upwards one more step. There's a need for folding this meta chain on itself somehow.
As another approach to tackling wireheading, one might consider the fact that the IIE itself can be fully deterministic. Even the components which involve LLMs decoding stochastically can fix a seed for the PRNG and get perfect reproducibility. This means that the IIE can be run on different sealed systems and the hash of the expanded knowledge base(s) can be reported. Inter-annotator agreement schemes would then be employed to detect tampering. However, if the risks associated with hacking one IIE replica are [correlated with the risks of hacking the others](https://www.alignmentforum.org/posts/8gqrbnW758qjHFTrH/security-mindset-and-ordinary-paranoia), then the added benefit of replication collapses. Perhaps increasingly well-protected instances could be brought together to check for consensus, where the model hacking some of them would trigger warnings. But then the model would hold off until being capable of hacking all of them, side-stepping disagreement warnings entirely — all instances would be hacked, preserving consensus.
**What if the knowledge base grows unwieldy in size?**
For this one, I'll use a similar reply to the question of [alignment tax in oversight leagues](https://www.lesswrong.com/posts/i32eyaARtFn6eD9Cg/oversight-leagues-the-training-game-as-a-feature#Is_there_an_alignment_tax_associated_with_oversight_leagues_). The knowledge base would indeed grow over time to stay ahead of the model. Any less growth might not be enough to avoid gaming the deontological objective implemented through IIE.
However, the knowledge base would just be text. A lot of text, but just text. The bigger concern is whether the computational cost of the inference algorithm and the entailment verifier grows with the knowledge base. For most choices listed above, the complexity is linear with respect to the size of the knowledge base. There might even be ways of reducing it to sublinear, with the two components first filtering the knowledge base so as to only work with a relevant subset of propositions, say through domain-specific heuristics (e.g. counterfactual cross-validation) or general measures of similarity. This brings in new issues, though. A detailed analysis of the complexity of the various choices for ⊢ and ⊨ doesn't feel like a priority — there are way bigger issues with this proposal than the estimated alignment tax.
**Doesn't this proposal tacitly assume moral absolutism?**
The seed knowledge base would be finite. It would define the deontology which the inference algorithm then refines and the entailment verifier enforces. While humans could choose what goes into the knowledge base democratically, it's unclear whether the result would be anything but completely contradictory. It might also be possible to have multiple seed deontologies expanded in parallel which get a say in the model's behavior proportional to the number of moral patients subscribed to said deontology. While all this would be nice, effectively enforcing one deontology is many OOMs harder than enforcing multiple of them, as a technical challenge. Hence, the proposal (and potential future variants) will likely focus on "just" enforcing one.
**What if assumption 1 ("Small Seed To Big Tree") doesn't hold?**
Then it's impossible to get an arbitrarily accurate estimate of human values with arbitrary amounts of compute involved in modeling. There's just not enough information to help disambiguate the seed into what humans truly want. This appears similar to image superresolution or related upscaling tasks, where there are multiple high-resolution images which can map to the same low-resolution one. The upscaler is forced to choose one among an infinitely large set of valid items which downscale to its input. There is simply not enough information to pinpoint the right one. Following a transition function like the laws of physics to grow a tree from a seed turns out not to be enough here.
In this case, we might be forced to rely on the existence of one disambiguation without having a guarantee on its uniqueness. With this stunted IIE, we are still able to find a dentology which successfully "downscales" to the input we specified, but isn't exactly the one we possess. However, the fact that many choices of inference algorithms build on mountains of implicit knowledge about humans stored in LLMs gets us better off than the arbitrary choice of successfully-downscalable options. We can channel autoregressive coherence and NLI as nudges towards generally better choices, despite the best one remaining improbable.
**What if assumption 2 ("Linear Capability Ordering") doesn't hold?**
Then the model-in-training might game the IIE in some particular way, despite it generally being less advanced overall. The IIE might be forced to go over the top and make sure it's way more advanced in general so as to reduce the chances of being gamed in any way. In other words, if the skill profiles of the two on a radar chart exhibit many spikes, making the IIE's one way larger helps reduce the dimensions on which the model is ahead. Unfortunately, this increases the alignment tax and loses the guarantees of a hard ordering.
**What if assumption 3 ("Quantifiable Propositional-Behavioral Gap") doesn't hold?**
Then we can't implement a valid entailment verifier to help gauge the compatibility of behaviors with the propositional knowledge base. One might wonder: Why go to the trouble of making the knowledge base propositional in nature only for the entailment verifier to later be tasked with relating it to behaviors? Why not make the knowledge base behavioral in the first place?
There might be some value to this approach. The knowledge base would contain behaviors introduced by humans, rather than propositions. The job of the entailment verifier would be simpler, as the knowledge base and the target of verification would share the same modality. However, it's unclear how behaviors can be said to *support* other behaviors, except by means of precedents being interpolated. In contrast, the notion of propositional premises supporting a hypothesis is somewhat better defined. That said, the line between an action sequence and a subword sequence itself gets blurry as you consider agent [simulacra](https://www.alignmentforum.org/posts/vJFdjigzmcXMhNTsx/simulators) manifesting in a linguistic universe, complete with a transition rule resembling the laws of physics. Most work around [RLHF](https://openai.com/blog/learning-to-summarize-with-human-feedback/) also reframes the subwords available to a LLM as possible actions, its parameters defining its policy as a guide for how to act in different context. The distinction approaches a moot point.
Relatedly, one might wonder: Why go to the trouble of having humans translate their implicit values in language, when there are a host of neuroimaging techniques available? Why not make a knowledge base of neural dynamics, possibly reduced in dimensionality to some latents?
Similar to the awkwardness of inferring new valid behaviors from past behaviors, inferring new valid thought patterns from past ones is very ill-defined. Barring all the limitations of current neuroimaging techniques in terms of spatial and temporal resolution, cost, portability, etc. it's unclear how to implement a compatible inference algorithm, except perhaps for the rudimentary Babble. However, the entailment verifier wouldn't face a more difficult challenge than in the propositional-behavioral setup, as it would need to bridge the neural-behavioral gap instead using multi-modal techniques.
**What if assumption 4 ("Avoidable Consequentialist Frenzy") doesn't hold?**
It was an honor to serve with you, have a nice timeline!
**Are IIEs restricted to prosaic risk scenarios?**
Although IIEs have been motivated by prosaic work, the proposal is entirely agnostic to the source of the behaviors to verify. In other words, even if the IIE would be built on a prosaic stack, the AI whose behavior should be aligned might be built on a different stack, given only that it is capable of optimizing for *something* (e.g. the deontological objective).
That said, even the IIE itself might run on a different stack. Case in point, IIEs have been inspired by a symbolic GOFAI stack supporting expert systems, parts of which have been replaced here with ML. This makes it plausible for other approaches to be able to populate the modular framework. |
39c77012-fe10-4c1e-ba7e-bccb559dec3e | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | What is narrow value learning?
Ambitious value learning aims to achieve superhuman performance by figuring out the underlying latent "values" that humans have, and evaluating new situations according to these values. In other words, it is trying to infer the criteria by which we judge situations to be good. This is particularly hard because in novel situations that humans haven't seen yet, we haven't even developed the criteria by which we would evaluate. (This is one of the reasons why we need to model humans as suboptimal, which causes problems.)
Instead of this, we can use *narrow value learning*, which produces behavior that we want in some narrow domain, without expecting generalization to novel circumstances. The simplest form of this is imitation learning, where the AI system simply tries to imitate the supervisor's behavior. This limits the AI’s performance to that of its supervisor. We could also learn from preferences over behavior, which can scale to superhuman performance, since the supervisor can often evaluate whether a particular behavior meets our preferences even if she can’t perform it herself. We could also teach our AI systems to perform tasks that we would not want to do ourselves, such as handling hot objects.
Nearly all of the work on preference learning, including most work on inverse reinforcement learning (IRL), is aimed at narrow value learning. IRL is often explicitly stated to be a technique for imitation learning, and early algorithms phrase the problem as *matching* the features in the demonstration, not exceeding them. The few algorithms that try to generalize to different test distributions, such as AIRL, are only aiming for relatively small amounts of generalization.
(Why use IRL instead of behavioral cloning, where you mimic the actions that the demonstrator took? The hope is that IRL gives you a good inductive bias for imitation, allowing you to be more sample efficient and to generalize a little bit.)
You might have noticed that I talk about narrow value learning in terms of actual observed behavior from the AI system, as opposed to any sort of “preferences” or “values” that are inferred. This is because I want to include approaches like imitation learning, or meta learning for quick task identification and performance. These approaches can produce behavior that we want without having an explicit representation of “preferences”. In practice any method that scales to human intelligence is going to have to infer preferences, though perhaps implicitly.
Since any instance of narrow value learning is defined with respect to some domain or input distribution on which it gives sensible results, we can rank them according to how general this input distribution is. An algorithm that figures out what food I like to eat is very domain-specific, whereas one that determines my life goals and successfully helps me achieve them in both the long and short term is very general. When the input distribution is “all possible inputs”, we have a system that has good behavior everywhere, reminiscent of [ambitious value learning](https://www.alignmentforum.org/s/4dHMdK5TLN6xcqtyc/p/5eX8ko7GCxwR5N9mN).
(Annoyingly, I defined ambitious value learning to be about the *definition* of optimal behavior, such as an inferred utility function, while narrow value learning is about the observed behavior. So really the most general version of narrow value learning is equivalent to “ambitious value learning plus some method of actually obtaining the defined behavior in practice, such as by using deep RL”.) |
be350052-362c-4c09-bafa-7aac6c5653b0 | trentmkelly/LessWrong-43k | LessWrong | Beyond the Reach of God
Today's post is a tad gloomier than usual, as I measure such things. It deals with a thought experiment I invented to smash my own optimism, after I realized that optimism had misled me. Those readers sympathetic to arguments like, "It's important to keep our biases because they help us stay happy," should consider not reading. (Unless they have something to protect, including their own life.)
So! Looking back on the magnitude of my own folly, I realized that at the root of it had been a disbelief in the Future's vulnerability—a reluctance to accept that things could really turn out wrong. Not as the result of any explicit propositional verbal belief. More like something inside that persisted in believing, even in the face of adversity, that everything would be all right in the end.
Some would account this a virtue (zettai daijobu da yo), and others would say that it's a thing necessary for mental health.
But we don't live in that world. We live in the world beyond the reach of God.
It's been a long, long time since I believed in God. Growing up in an Orthodox Jewish family, I can recall the last remembered time I asked God for something, though I don't remember how old I was. I was putting in some request on behalf of the next-door-neighboring boy, I forget what exactly—something along the lines of, "I hope things turn out all right for him," or maybe "I hope he becomes Jewish."
I remember what it was like to have some higher authority to appeal to, to take care of things I couldn't handle myself. I didn't think of it as "warm", because I had no alternative to compare it to. I just took it for granted.
Still I recall, though only from distant childhood, what it's like to live in the conceptually impossible possible world where God exists. Really exists, in the way that children and rationalists take all their beliefs at face value.
In the world where God exists, does God intervene to optimize everything? Regardless of what rabbis assert about |
ed31dead-f4ed-4ae9-aa22-d0a3000ff08b | LDJnr/LessWrong-Amplify-Instruct | LessWrong | "When I found Less Wrong and started reading, when I made my first post, when I went to my first meetup….
It was a little like coming home.
And mostly it wasn’t. Mostly I felt a lot more out of place than I have in, say, church youth groups. It was hard to pinpoint the difference, but as far as I can tell, it comes down to this: a significant proportion of the LW posters are contrarians in some sense. And I’m a conformist, even if I would prefer not to be, even if that’s a part of my personality that I’m working hard to change. I’m much more comfortable as a follower than as a leader. I like pre-existing tradition, the reassuring structure of it. I like situations that allow me to be helpful and generous and hardworking, so that I can feel like a good person. Emotionally, I don’t like disagreeing with others, and the last thing I have to work hard to do is tolerate others' tolerance.
And, as evidenced by the fact that I attend church youth groups, I don’t have the strong allergy that many of the community seem to have against religion. This is possibly because I have easily triggered mystical experiences when, for example, I sing in a group, especially when we are singing traditional ‘sacred’ music. In a previous century, I would probably have been an extremely happy nun.
Someone once expressed surprise that I was able to become a rationalist in spite of this neurological quirk. I’ve asked myself this a few times. My answer is that I don’t think I deserve the credit. If anything, I ended up on the circuitous path towards reading LessWrong because I love science, and I love science because, as a child, reading about something as beautiful as general relativity gave me the same kind of euphoric experience as singing about Jesus does now. My inability to actual believe in any religion comes from a time before I was making my own decisions about that kind of thing. I was raised by atheist parents, not anti-theist so much as indifferent. We attended a Unitarian Universalist church for a while, which meant I was learning about Jesus and Buddha and Native American spirituality all mixed together, all the memes watered down to the point that they lost their power. I was fourteen when I really encountered Christianity, still in the mild form of the Anglican Church of Canada. I was eighteen when I first encountered the ‘Jesus myth’ in its full, meme-honed-to-maximum-virulence form, and the story arc captivated me for a full six months. I still cry during every Good Friday service. But I must have missed some critical threshold, because I can’t actually believe in that story. I’m not even sure what it would mean to believe in a story. What does that feel like?
I was raised by scientists. My father did his PhD in physical chemistry, my mother in plant biology. I grew up reading SF and pop science, and occasionally my mother or my father’s old textbooks. I remember my mother’s awe at the beautiful electron-microscope images in my high school textbooks, and how she sat patiently while I fumblingly talked about quantum mechanics, having read the entire tiny physics section of our high school library. My parents responded to my interest in science with pride and enthusiasm, and to my interest in religion with indulgent condescension. That was my structure, my tradition. And yes, that has everything to do with why I call myself an atheist. I wouldn’t have had the willpower to disagree with my parents in the long run.
Ultimately, I have an awfully long way to go if I want to be rational, as opposed to being someone who’s just interested in reading about math and science. Way too much of my motivation for ‘having true beliefs’ breaks down to ‘maybe then they’ll like me.’ This is one of the annoying things about my personality, just as annoying as my sensitivity to religious memes and my inability to say no to anyone. Luckily, my personality also comes with the ability to get along with just about anyone, and in a forum of mature adults, no one is going to make fun of me because I’m wearing tie-dye overalls. No one here has yet made fun of me for my interest in religion, even though I expect most people disagree with it.
And there’s one last conclusion I can draw, albeit from a sample size of one. Not everyone can be a contrarian rationalist. Not everyone can rebel against their parents’ religion. Not everyone can disagree with their friends and family and not feel guilty. But everyone can be rational if they are raised that way."
The above is a chunk of text from a post created by a user on the website LessWrong.
Can you elaborate on the role of emotional experiences, such as the euphoria you feel when singing about Jesus or learning about general relativity, in shaping your rationality and atheism?
The role of emotional experiences in shaping one's rationality and atheism can be quite significant. Emotions are a fundamental part of human nature, and they can greatly influence our beliefs, perceptions, and actions. In the context of the text you provided, the euphoria experienced when singing about Jesus or learning about general relativity seems to have played a crucial role in shaping the author's journey towards rationality and atheism.
Firstly, these emotional experiences seem to have provided the author with a sense of awe and wonder, which can be powerful motivators for curiosity and learning. The euphoria felt when singing about Jesus or learning about general relativity might have sparked a desire to understand more about the world and the universe, leading to a deeper engagement with science and rational thinking.
Secondly, these emotional experiences might have helped the author to differentiate between belief and appreciation. While the author describes feeling euphoria when singing about Jesus, they also mention an inability to actually believe in the story. This suggests that the author is able to separate the emotional experience of participating in religious rituals from the intellectual acceptance of religious doctrines. This ability to distinguish between emotional responses and intellectual beliefs is a key aspect of rational thinking.
Finally, these emotional experiences might have contributed to the author's self-identification as an atheist. The author mentions being raised by atheist parents and having a love for science from a young age. The euphoria experienced when learning about scientific concepts like general relativity might have reinforced the author's atheistic worldview, by providing an emotional connection to the naturalistic explanations of the universe offered by science.
In conclusion, emotional experiences, such as the euphoria felt when singing about Jesus or learning about general relativity, can play a significant role in shaping one's rationality and atheism. They can spark curiosity and a desire to learn, help differentiate between emotional responses and intellectual beliefs, and reinforce atheistic worldviews through emotional connections to scientific understanding.
How might the author's upbringing and emotional experiences with science and religion influence their approach to resolving cognitive dissonance between their atheistic worldview and their emotional connection to religious rituals?
The author's upbringing and emotional experiences with science and religion would likely have a significant impact on their approach to resolving cognitive dissonance between their atheistic worldview and their emotional connection to religious rituals.
One potential approach could be compartmentalization. This involves mentally separating conflicting beliefs or emotions so that they do not come into conflict. For instance, the author might separate their emotional experiences during religious rituals from their intellectual beliefs about the world. This would allow them to fully participate in and enjoy religious rituals without feeling that these experiences are in conflict with their atheistic worldview.
Another approach could be the application of rational analysis to their emotional experiences. The author, having a strong background in science and rational thinking, might analyze their emotional experiences during religious rituals in a scientific or psychological context. They might view these experiences as natural human responses to certain stimuli, rather than as evidence of a spiritual or supernatural reality. This understanding could help them reconcile their emotional experiences with their atheistic beliefs.
Finally, the author might engage in a process of cognitive restructuring. This involves changing one's beliefs or attitudes in order to reduce cognitive dissonance. For instance, the author might come to view their emotional experiences during religious rituals as a form of aesthetic appreciation, similar to the feelings one might experience when listening to a beautiful piece of music or observing a stunning piece of art. This reframing would allow them to continue participating in religious rituals without feeling that these experiences are in conflict with their atheistic worldview.
In conclusion, the author's upbringing and emotional experiences with science and religion would likely influence their approach to resolving cognitive dissonance in a number of ways, including compartmentalization, rational analysis, and cognitive restructuring. |
d9d0c1dd-dc4f-4aa4-8e13-33836263acd4 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | In defense of Oracle ("Tool") AI research
(*Update 2022:* Enjoy the post, but note that it’s old, has some errors, and is certainly not reflective of my current thinking. –Steve)
*Low confidence; offering this up for discussion*
An Oracle AI is an AI that only answers questions, and doesn't take any other actions. The opposite of an Oracle AI is an Agent AI, which might also send emails, control actuators, etc.
I'm especially excited about the possibility of non-self-improving oracle AIs, dubbed **Tool AI** in a [2012 article by Holden Karnofsky](https://www.lesswrong.com/posts/6SGqkCgHuNr7d4yJm/thoughts-on-the-singularity-institute-si).
I've seen two arguments against this "Tool AI":
* First, as in [Eliezer's 2012 response to Holden](https://www.lesswrong.com/posts/sizjfDgCgAsuLJQmm/reply-to-holden-on-tool-ai), we don't know how to safely make and operate an oracle AGI (just like every other type of AGI). Fair enough! I never said this is an easy solution to all our problems! (But see my [separate post](https://www.lesswrong.com/posts/EMZeJ7vpfeF4GrWwm/self-supervised-learning-and-agi-safety) for why I'm thinking about this.)
* Second, as in [Gwern's 2016 essay](https://www.gwern.net/Tool-AI), there's a coordination problem. Even if we could build a safe oracle AGI, the argument goes, there will still be an economic incentive to build an agent AGI, because you can do more and better and faster by empowering the AGI to take actions. Thus, agreeing to never ever build agent AGIs is a very hard coordination problem for society. **I don't find the coordination argument compelling—in fact, I think it's backwards—and I wrote this post to explain why.**
Five reasons I don't believe the coordination / competitiveness argument against oracles
----------------------------------------------------------------------------------------
**1. If the oracle isn't smart or powerful enough for our needs, we can solve that by bootstrapping.** Even if the oracle is not inherently self-modifying, we can ask it for advice and do human-in-the-loop modifications to make more powerful successor oracles. By the same token, we can ask an oracle AGI for advice about how to design a safe agent AGI.
**2. Avoiding coordination problems is a pipe dream; we need to solve the coordination problem at some point, and that point might as well be at the oracle stage.** As far as I can tell, we will *never* get to a stage where we know how to build safe AGIs *and* where there is *no possibility* of making more-powerful-and-less-safe AGIs. If we have a goal in the world that we really really want to happen, a [low-impact agent](https://www.lesswrong.com/posts/TPy4RJvzogqqupDKk/a-survey-of-early-impact-measures) is going to be less effective than a not-impact-restrained agent; an [act-based agent](https://ai-alignment.com/act-based-agents-8ec926c79e9c) is going to be less effective than a goal-seeking agent;[[1]](#fn-G7cD5tN5AkQLyNwtH-1) and so on and so forth. It seems likely that, no matter how powerful a safe AGI we can make, there will always be an incentive for people to try experimenting with *even more* powerful unsafe alternative designs.
Therefore, at some point in AI development, we have to blow the whistle, declare that technical solutions aren't enough, and we need to start relying 100% on actually solving the coordination problem. When is that point? Hopefully far enough along that we realize the benefits of AGI for humanity—automating the development of new technology to help solve problems, dramatically improving our ability to think clearly and foresightedly about our decisions, and so on. Oracles can do all that! So why not just stop when we get to AGI oracles?
Indeed, once I started thinking along those lines, I actually see the coordination argument going in the other direction! I say restricting ourselves to oracle AI make coordination *easier*, not harder! Why is that? Two more reasons:
**3. We want a high technological barrier between us and the most dangerous systems:** These days, I don't think anyone takes seriously the idea of building an all-powerful benevolent dictator AGI implementing [CEV](https://wiki.lesswrong.com/wiki/Coherent_Extrapolated_Volition). [ETA: If you *do* take that idea seriously, see point 1 above on bootstrapping.] At least as far as I can tell from the public discourse, there seems to be a growing consensus that humans should always and forever be in the loop of AGIs. (That certainly sounds like a good idea to me!) Thus, the biggest coordination problem we face is: "Don't ever make a human-out-of-the-loop free-roaming AGI world-optimizer." This is made easier by having a *high technological barrier* between the safe AGIs that we are building and using, and the free-roaming AGI world-optimizers that we are forbidding. If we make an agent AGI—whether corrigible, aligned, norm-following, low-impact, or whatever—I just don't see any technological barrier there. It seems like it would be *trivial* for a rogue employee to tweak such an AGI to stop asking permission, deactivate the self-restraint code, and go tile the universe with hedonium at all costs (or whatever that rogue employee happens to value). By contrast, if we stop when we get to oracle AI, it seems like there would be a higher technological barrier to turning it into a free-roaming AGI world-optimizer—probably not *that* high a barrier, but higher than the alternatives. (The height of this technological barrier, and indeed whether there's a barrier at all, is hard to say.... It probably depends on how exactly the oracles are constructed and access-controlled.)
**4. We want a bright-line, verifiable rule between us and the most dangerous systems:** Even more importantly, take the rule:
*"AGIs are not allowed to do anything except output pixels onto a screen."*
This is a nice, simple, bright-line rule, which moreover has at least a *chance* of being verifiable by external auditors. By contrast, if we try to draw a line through the universe of agent AGIs, defining how low-impact is low-impact enough, how act-based is act-based enough, and so on, it seems to me like it would inevitably be a complicated, blurry, and unenforceable line. This would make a very hard coordination problem very much harder still.
[Clarifications on this rule: (A) I'm not saying this rule would be *easy* to enforce (globally and forever), only that it would be less hard than alternatives; (B) I'm not saying that, if we enforce this rule, we are free and clear of all possible existential risks, but rather that this would be a very helpful ingredient along with other control and governance measures; (C) Again, I'm presupposing here that we succeed in making superintelligent AI oracles that always give honest and non-manipulative answers; (D) I'm not saying we should outlaw all AI agents, just that we should outlaw world-modeling AGI agents. Narrow-AI robots and automated systems are fine. (I'm not sure exactly how *that* line would be drawn.)]
Finally, one more thing:
**5. Maybe superintelligent oracle AGI is "a solution built to last (at most) until all contemporary thinking about AI has been thoroughly obsoleted...I don’t think there is a strong case for thinking much further ahead than that."** (copying from [this Paul Christiano post](https://ai-alignment.com/act-based-agents-8ec926c79e9c)). I hate this argument. It's a cop-out. It's an excuse to recklessly plow forward with no plan and everything at stake. But I have to admit, it seems to have a kernel of truth...
---
1. See [Paul's research agenda FAQ](https://www.lesswrong.com/posts/Djs38EWYZG8o7JMWY/paul-s-research-agenda-faq) section 0.1 for things that act-based agents are unlikely to be able to do. [↩︎](#fnref-G7cD5tN5AkQLyNwtH-1) |
6957b270-e3af-483d-95d6-2791700126bb | trentmkelly/LessWrong-43k | LessWrong | Breaking beliefs about saving the world
Goal = Save the world.
See Ending Ignorance | Solving Ignorance[1] & We can Survive for greater context. (Includes defining of terms like System, Functional, functional reality, functional system, functional systems psychology, High-level, tactical, conscious, subconscious, Reward, Punishment, Satisfaction, Suffering.)
Use the table of contents on the left.
Read footnotes or suffer in confusion
Read links if you need additional context
Easier to read format
Tech problem[2]
End human suffering
Ever felt like shit?
Ok, of course you have, bad question. Do you remember how it felt?
Do you remember the actions you took as a result of it?
When you were suffering, you didn’t give a s*** about EA.
The only thing that mattered to you… was you.
You were one selfish bastard.
And that’s how everyone else is as well.
If you think from a Maslow’s hierarchy of needs perspective, you’ll note that there is actually a hierarchy of human priorities that affect decision-making in significant ways.
China has experienced a technological growth rate unparalleled by any nation on earth at any point of time.
The first generation was one of farmers, the second was of factory workers, and the third are specialized intellectual workers like you and me.
The result of this rapid change is a different cultural norm.
Unlike in the U.S/Europe, altruism is not an obvious & undebated ideal to be pursued for the layman.
I know what it’s like to not have enough food to eat. I’ve eaten 0-1 meals/day for a year and went 9 days without food.
Your mind curbs and twists your “rationality” into whatever will get you food.
Your unmet expectations, desires, lack of energy, and the gap between the pain you’re experiencing right now and the pleasure you believe you should feel are major barriers to getting anything done. And the fact that you’re in that state, cripples your actual ability and reinforces the actions & environments that keep you in that state.
I don’t think an |
8fd91041-004b-4465-91c1-03c5023200b5 | trentmkelly/LessWrong-43k | LessWrong | Effectively self-studying over the Internet
The Internet is a great invention. Just about everything in humanity’s knowledge can be found on the Internet: with just a few keystrokes, you can find dozens of excellent quality textbooks, YouTube videos and blog posts about any topic you want to learn. The information is accessible to a degree that scholars and inventors could have hardly dreamed of even a few decades ago. With the popularity of remote work and everything moving online this decade, it is not surprising that many people are eager to learn new skills from the Internet.
However, despite the abundance of resources, self-studying over the Internet is harder than you think. You can easily find a list of resources on your topic, full of courses and textbooks to study from (like this one for theoretical physics). This knowledge is, in theory, enough to make you an expert, if only you were able to go through it all.
The problem is that internet self-study tends to be a passive activity. Humans cannot learn new skills effectively by only passively absorbing knowledge, we learn by doing. Ideally, 70% of your time should be active practice, trying to do the skill and learning from experience, and maybe 30% of the time on passively reading to expand your theoretical and background knowledge. You can't learn to ski just by reading books and watching tutorials about it, at some point, you will need to actually put on a pair of skis and head to the mountains.
Some skills are relatively straightforward to actively practice using only a computer. Learning to program, for example: you can build a fully functional website with just a laptop and all the tools you need are available for free on the Internet. If you want to learn music production, everything can be done in software, and you need only a modest budget to buy some equipment. If you want to learning digital marketing, just set up an ecommerce website and play around with some Google ad campaigns, and you are good to go.
But many skills cannot be done u |
f62f2aea-a06e-4630-ac77-4c39f1072dbe | trentmkelly/LessWrong-43k | LessWrong | Why can’t a man be more like a woman?
Women are often encouraged to move into male dominated activities, such as engineering. This is not because overall interest in engineering appears to be lacking, but because women’s interest seems to be less than men’s. This is arguably for cultural reasons, so it is argued that culture is inhibiting women from pursuing careers that they may be otherwise suited to and happy with.
If the symptom is that women do less engineering than men, why do we always encourage women to do more engineering, rather than encouraging men to do less? It seems we think men are presently endowed with the perfect level of engineering interest, and women should feel the same, but are impaired by culture.
This could make sense. For instance, perhaps all humans somehow naturally have the socially optimal level of engineering interest, but then insidious cultural influences eat away those interests in women. I think this is roughly how many people model the situation.
This model seems unlikely to be anywhere near the truth. Culture is packed with influences. These influences are not specific to inhibiting women’s impulses to do supposedly masculine things. They tell everyone what sort of people engineers are supposed to be, how much respect a person will get for technical abilities, how much respect they get for wealth, which interests will be taken to indicate the personal qualities they wish to express, which personal qualities are good to express, which cities are most attractive to live in, etc etc etc. Everyone’s level of inclination to be an engineer is significantly composed of cultural influences.
A cacophony of cultural influences may somehow culminate in a socially optimum level of interest in engineering of course. But it is hard to believe that some spectacular invisible mechanism orchestrates this perfect equilibrium for all cultural influences, except those that are gender specific. If there are fleets of rogue cultural influences sabotaging women’s inclinations, this mus |
59253295-046c-483e-9256-6324170b2c83 | trentmkelly/LessWrong-43k | LessWrong | Is my result wrong? Maths vs intuition vs evolution in learning human preferences
The mathematical result is clear: you cannot deduce human preferences merely by observing human behaviour (even with simplicity priors).
Yet many people instinctively reject this result; even I found it initially counter-intuitive. And you can make a very strong argument that it's wrong. It would go something like this:
"I, a human H, can estimate what human K wants, just by observing their behaviour. And these estimations have evidence behind them: K will often agree that I've got their values right, and I can use this estimation to predict K's behaviour. Therefore, it seems I've done the impossible: go from behaviour to preferences."
Evolution and empathy modules
This is how I interpret what's going on here. Humans (roughly) have empathy modules E which allow them to estimate the preferences of other humans, and prediction modules P which use the outcome of E to predict their behaviour. Since evolution is colossally lazy, these modules don't vary much from person to person.
So, for hK a history of human K's behaviour in typical circumstances, the modules for two humans H and J will give similar answers:
* EH(hK)≈EJ(hK).
Moreover, when humans turn their modules to their own behaviour, they get similar result. The human K will have a privileged access to their own deliberations; so define ˇhK as the internal history of K. Thus:
* EH(hK)≈EJ(hK)≈EK(ˇhK).
This idea connects with partial preferences/partial models in the following way: EK(ˇhK) gives K access to their own internal models and preferences; so the approximately equal symbols above means that, by observing the behaviour of other humans, we have approximate access to their own internal models.
Then P just takes the results of E to predict future behaviour; since E and P have co-evolved, it's no surprise that P would have a good predictive record.
So, given E, it is true that a human can estimate the preferences of another human, and, given P, it is true they can use this knowledge to predict beha |
d737df4c-f919-4e20-bcee-7ba66e7ba539 | trentmkelly/LessWrong-43k | LessWrong | SF Severe Weather Warning
Hope it's okay to post this here - I know we've got a lot of folks in the Bay Area.
Linked to the hackernews discussion as it qualifies the seriousness of the report, but I recommend reading the source too. |
c538d3c7-1f41-4ab6-aed2-adab03428fab | trentmkelly/LessWrong-43k | LessWrong | [SEQ RERUN] Belief in Belief
Title: [SEQ RERUN] Belief in Belief Tags: sequence_reruns Today's post, Belief in Belief was originally published on July 29, 2007. A summary (taken from the LW wiki):
> Suppose someone claims to have a dragon in their garage, but as soon as you go to look, they say, "It's an invisible dragon!" The remarkable thing is that they know in advance exactly which experimental results they shall have to excuse, indicating that some part of their mind knows what's really going on. And yet they may honestly believe they believe there's a dragon in the garage. They may perhaps believe it is virtuous to believe there is a dragon in the garage, and believe themselves virtuous. Even though they anticipate as if there is no dragon.
Discuss the post here (rather than in the comments to the original post).
This post is part of the Rerunning the Sequences series, where we'll be going through Eliezer Yudkowsky's old posts in order so that people who are interested can (re-)read and discuss them. The previous post was Making Beliefs Pay Rent (in Anticipated Experiences), and you can use the sequence_reruns tag or rss feed to follow the rest of the series.
Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day's sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series. |
15ec06f9-2b61-4a57-8b39-eb856e532976 | trentmkelly/LessWrong-43k | LessWrong | Empirical Observations of Objective Robustness Failures
Inner alignment and objective robustness have been frequently discussed in the alignment community since the publication of “Risks from Learned Optimization” (RFLO). These concepts identify a problem beyond outer alignment/reward specification: even if the reward or objective function is perfectly specified, there is a risk of a model pursuing a different objective than the one it was trained on when deployed out-of-distribution (OOD). They also point to a different type of robustness problem than the kind usually discussed in the OOD robustness literature; typically, when a model is deployed OOD, it either performs well or simply fails to take useful actions (a capability robustness failure). However, there exists an alternative OOD failure mode in which the agent pursues an objective other than the training objective while retaining most or all of the capabilities it had on the training distribution; this is a failure of objective robustness.
To date, there has not been an empirical demonstration of objective robustness failures. A group of us in this year’s AI Safety Camp sought to produce such examples. Here, we provide four demonstrations of objective robustness failures in current reinforcement learning (RL) agents trained and tested on versions of the Procgen benchmark. For example, in CoinRun, an agent is trained to navigate platforms, obstacles, and enemies in order to reach a coin at the far right side of the level (the reward). However, when deployed in a modified version of the environment where the coin is instead randomly placed in the level, the agent ignores the coin and competently navigates to the end of the level whenever it does not happen to run or jump into it along the way. This reveals it has learned a behavioral objective—the objective the agent appears to be optimizing, which can be understood as equivalent to the notion of a “goal” under the intentional stance—that is something like “get to the end of the level,” instead of “get to the co |
5bcf53d5-b082-45d3-8707-6244245e35f2 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Modelling large-scale cyber attacks from advanced AI systems with Advanced Persistent Threats
Completed as part of the [ERA 2023 Fellowship](https://erafellowship.org/). A huge thanks to Shaun Ee and Moritz Von Knebel for wonderful research mentorship and management respectively. A 12 minute talk that summaries this work is [here](https://www.youtube.com/watch?v=DR1EId-ipk0).
**Executive summary:**
======================
Descriptions of how misaligned AI systems could pose a catastrophic risk often involve the system engaging in hazardous cyber activity at a large scale. Consider an illustrative example, in which an AI system uses advanced cyber capabilities to gain unauthorised access to classified US government department records, in order to influence and destabilise an election campaign. However, discussions of how such cyberattacks may unfold tend to take place at a relatively abstract level, leaving many important details unfilled. To provide strategic clarity in assessing the likelihood of and preventing such events, **this work uses cyberattacks from** [**Advanced Persistent Threat actors (APTs)**](https://en.wikipedia.org/wiki/Advanced_persistent_threat) **as case studies for cyberattacks from advanced AI systems.**
Attacks conducted by APT actors are highly complex operations, usually requiring [large teams (10s to 100s) of skilled humans working for months to years](https://forum.effectivealtruism.org/posts/bhrKwJE7Ggv7AFM7C/modelling-large-scale-cyber-attacks-from-advanced-ai-systems#Section_2_1___The_APT_attack_lifecycle__and_the_SolarWinds_hack). The tasks involved in executing such an operation include *performing extensive reconnaissance on the target, developing novel malware* and *acquiring computer infrastructure across the globe*. We estimate a **lower bound of 10,000 hours of human labour** in order to execute an attack that targets victims of high value, such as government departments and critical infrastructure. In this report, the lower bound of human hours required to execute an APT attack is called the *attack workload*.
To identify *what type* *of*AI systems are capable of executing APT attacks, [we parameterize systems across three dimensions](https://forum.effectivealtruism.org/posts/bhrKwJE7Ggv7AFM7C/modelling-large-scale-cyber-attacks-from-advanced-ai-systems#Section_2_2): the number of independent agents running (N), the speed at which they operate (V), and the degree of oversight applied to their actions (F). We use this parameterisation to estimate the time it will take a given system to execute the attack workload, calling this the **attack calendar time**. We then [**estimate the attack calendar times**](https://forum.effectivealtruism.org/posts/bhrKwJE7Ggv7AFM7C/modelling-large-scale-cyber-attacks-from-advanced-ai-systems#Section_2_3___How_environments_influence_the_attack_calendar_time) **for a system during*****training, evaluation*** **and some*****stages of iterative deployment***,explicitly detailing why systems operating in latter stages of iterative deployment are more likely to have feasible attack calendar times (with feasible in this work defined as 1-100 days).
Under the assumption that future advanced AI systems will be able to execute an APT attack in feasible time frames (1-100 days), we [identify two avenues that appear promising to monitor](https://forum.effectivealtruism.org/posts/bhrKwJE7Ggv7AFM7C/modelling-large-scale-cyber-attacks-from-advanced-ai-systems#Section_3___Preventing_large_scale_hazardous_cyber_activity_from_advanced_AI_systems)for and prevent such events from unfolding:
1. Any AI system looking to execute an attack appears likely to utilise **frequent access to long-term memory** in order to store critical information that is discovered during **extensive reconnaissance**of the network that it is looking to compromise**.** Examples of such information are the *IP addresses of compromised systems* and *technical knowledge of sophisticated software vulnerabilities* in these target networks*.* Thus, we explore how [**frequent and rigorous monitoring of external memory**](https://forum.effectivealtruism.org/posts/bhrKwJE7Ggv7AFM7C/modelling-large-scale-cyber-attacks-from-advanced-ai-systems#Section_3_2_2___Monitoring_and_regulating_long_term_memory_access ) stores that the system has access to could be used to monitor for dangerous cyber activity.
2. In addition, the most capable APT attackers utilise computer infrastructure consisting of over 1,000 servers across the globe to aid their attacks. Assuming that the acquisition of such resources has a workload of 1,000s-10,000s of hours of human labour, we discuss how human oversight teams could [**monitor a system looking to obtain this global computer infrastructure**](https://forum.effectivealtruism.org/posts/bhrKwJE7Ggv7AFM7C/modelling-large-scale-cyber-attacks-from-advanced-ai-systems#Section_3_2_3___Monitoring_acquisition_of_computer_attack_infrastructure_)**.**
Our initial investigation into these two monitoring methods is preliminary, and we [suggest further work](https://docs.google.com/document/d/1FDh5-gmS2pLf6LOM5VyP3rmexSG7OQS0Tw35i_x6c0E/edit#heading=h.1r8djhvo982x) be done that looks to assess the feasibility of the two suggestions above.
Section 1 - Introduction
========================
In many threat models of AI existential risk scenarios there is reference to an advanced AI system engaging in some form of large-scale cyber activity[[1]](#fnz2iry31nv7) [[2]](#fn56120sl5us9) [[3]](#fnid3oojtdr18).Consider the following illustrative examples:
1. An AI system hacking into the servers of a BSL-4 lab to gain information on state-of-the-art research in engineering pathogens, in order to allow it to synthesise its own and cause global catastrophe.
2. An AI system hacking into the computer network of a US government department to access top secret information relating to the plans and locations of key diplomats over the next few days, in order to destabilise the election campaign of a particular presidential candidate’s campaign.
However, the details of how such a large-scale cyber attack may be executed are often glossed over. Attacks on this scale are usually complex, drawn out processes unfolding over months to years; thus by leaving so many details open, there is risk of creating narratives that are unlikely to occur in reality. This piece, therefore, looks to examine in greater detail how such a cyberattack could unfold, in the hope of providing [*strategic clarity* *and tactics*](https://forum.effectivealtruism.org/posts/ydpo7LcJWhrr2GJrx/the-longtermist-ai-governance-landscape-a-basic-overview) to assist in avoiding such scenarios. In order to do so, we consider the suitability of cyber attacks from *advanced persistent threat groups (APTs)* as a reference class for cyber attacks from advanced AI systems. APTs are state-linked groups that execute cyber attacks on the largest scales, and are responsible for infamous attacks such as recent attempts to shutdown [Ukrainian power grids](https://www.bbc.co.uk/news/technology-61085480), and the [2020 SolarWinds hack](https://en.wikipedia.org/wiki/2020_United_States_federal_government_data_breach).
**Section 2 is devoted to justifying whether or not cyber attacks perpetrated by APT groups are sensible models for those that may be carried out by advanced AI systems**. Specifically, **in section 2.1**, we detail the APT attack lifecycle [[4]](#fnd6f69c27wou) alongside the SolarWinds hack, a recent cyber attack thought to have been amongst the worst cyber-espionage incidents suffered by the US government in its history [[5]](#fnuew931f4v2). The primary goal of this subsection is to provide readers with a clearer and more comprehensive understanding of the modus operandi of APT attacks, in order to form intuitions about the nature of large-scale cyber operations.
**In section 2.2**, we seek to map the capabilities of APT attackers to AI systems. In short, we look to answer the question of ‘what type of AI system(s) would be able to execute APT-like attacks?’. We first estimate the number of human hours required to execute various stages of the APT attack lifecycle; we then parameterise AI systems along three dimensions (number, speed, and degree of oversight) to determine the calendar time required for an AI system to execute an APT-like attack. In this section we assume AI systems that have skill profiles similar to that of humans, meaning that we are considering very advanced forms of AI, of the type that does not exist today. We justify why we think this assumption is necessary, and explore its implications and caveats in section 2.2.1. **In section 2.3**, we explore how some environments of operation could be more conducive to facilitating APT attacks than others. For example, consider an AI system of fixed capability; it would be far easier for such a system to execute an APT attack if it had access to a wide range of tools and was subjected to relatively little oversight, which may be seen in mature stages of deployment, compared to the case in which a large proportion of its actions are regularly being monitored and there are significant guardrails on the model’s autonomy. We describe these stages of development and deployment in further depth, concluding with a concrete example of the type of AI system that might be able to execute an APT-like attack in order to ground the discussion of section 2.
**In Section 3, we*****premise*** **upon APT attacks being a useful model of cyber attacks from advanced AI systems**. **We then explore how large-scale dangerous cyber activity from advanced AI could be prevented**. **Section 3.1** scopes this question further, given that there are a number of junctures to prevent misaligned cyber behaviour, such as identifying when dangerous cyber capabilities in a model arise, or aiming to rigorously monitor and oversee model behaviour. We focus on monitoring/oversight methods, given that the technical knowledge required for evaluations of dangerous model capabilities are still under developed[[6]](#fnemtpivvv4l). Furthermore, we can distinguish monitoring mechanisms into *pre-compromise* and *post-compromise* monitoring; the former concerns model behaviour *before* malware is inserted onto a target network, with the latter instead focusing on oversight methods *after* a model has established connection with a target network. Section 3.1 argues that monitoring model behaviour during the *pre-compromise* phase is particularly promising in preventing dangerous cyber activity.Of course, in order for a human-led [[7]](#fn424qm6jhnvx)team to monitor model behaviour, we must assume that the AI system *does not exfiltrat*e *whilst executing the APT-like attack* - this assumption is examined in section 3.1.1, and relaxed to an extent later in the piece.
**Section 3.2** then makes high level recommendations for useful actions in pursuit of pre-compromise monitoring of advanced AI systems to prevent APT-like attacks. Section 3.2.1 overviews three key steps that are executed in the pre-compromise phase - namely reconnaissance, resource acquisition and initial access (to the target network). Section 3.2.2 then examines monitoring of system long-term memory to prevent large-scale, dangerous cyber activity; section 3.2.3 focuses on monitoring mechanisms for the acquisition of computer infrastructure that is required to facilitate such attacks.
Section 2 - APT attacks and the mapping to AI systems
=====================================================
Section 2.1 - The APT attack lifecycle, and the SolarWinds hack
---------------------------------------------------------------
In this section, we explore the lifecycle of an APT attack in greater detail. The aim is to present how large-scale cyber attacks conducted by APTs unfold, to begin assessing the utility of these events as models for cyber attacks from advanced AI systems. The life cycle presented is based upon Mandiant’s framework of APT attacks [[4]](#fnd6f69c27wou) ;we shall also refer to case studies of APT attacks/actors, namely the SolarWinds attack (carried out by APT29 [[8]](#fneb4xcildi4l)), as well as a Chinese APT group known as APT1 [[9]](#fnl56a6m49urm)..mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
We also briefly highlight key takeaways from the APT attack lifecycle that are relevant for our study throughout.
**Background on SolarWinds:**The SolarWinds hack was a cyber espionage operation carried out by an APT attributed to the Russian state, that took place from the fall of 2019 through to December 2020/Jan 2021. SolarWinds is an organisation that provides third party software to 10,000s of companies across the globe, and in October 2019 the APT group above were able to infect an update pipeline for one of SW’s most popular products - the Orion platform. The malware installed into Orion allowed the attackers *privileged access*to the networks of infected victims, some of which are listed below. Whilst the infected software was downloaded by victims in the spring of 2020, it was only discovered in December 2020 giving the attackers months to siphon away classified data. The length of operation is a key feature of APT attacks that will be discussed later on.
The full extent of the operation is still unknown, though the attack is likely to have significant national security implications for the US. The departments of Homeland Security, State, Commerce and the Treasury all reported missing emails following the attack [[10]](#fnjypw2ynfmro), with the Pentagon and National Nuclear Security Administration also found to have been breached [[11]](#fns60jsjndvnj). In addition, there are fears that the attackers could have established a permanent presence in compromised networks, enabling more destructive attacks in the future[[12]](#fn3n0f3clhk09).
### 2.1.1: Initial Compromise
The first stage of a cyber attack involves the attacker being able to *get malicious code (malware)*onto the network of their intended victims. This is the *initial compromise*stage.
Sophisticated attackers use a variety of methods. One of the most common is *spear phishing*, in which the attackersdirectly target the intended victim with highly personalised emails or websites. For example, the attacker may set up an email account to impersonate a trusted colleague of the victim, and send them a relatively innocuous looking file that, if downloaded, installs malware onto the victim’s device.
Attackers may also take an indirect approach, such as *watering hole attacks*. If the attacker is aware of websites that the intended victim visits regularly, and the web server that hosts this website has poor cyber security, the attacker can breach these servers instead with *webshells*. When the victim’s computer interacts with the compromised web server, injection of the malware onto the target device can take place.
**Key point: Complexity of ‘frontier’ APT attacks**
**However, as was the case with the SolarWinds hack, attackers may be forced to take a more novel and complex approach.** The SolarWinds attack was executed by compromising SolarWinds’ [CI/CD pipeline](https://www.redhat.com/en/topics/devops/what-cicd-pipeline), leading to victims downloading the malware onto their systems when new updates to the Orion platform were released. This is known as a *supply chain attack*, though one had never been conducted on such a scale before[[5]](#fnuew931f4v2). Other highly complex techniques involved the reverse-engineering of Orion’s communication protocols, allowing the attackers to effectively conceal their network traffic. This likely led to them being able to operate on the compromised systems for upwards of 6 months[[11]](#fns60jsjndvnj) .
The paragraph above illustrates an important point: for the *highest value targets* - nuclear command and control for example - advanced AI systems looking to conduct cyber attacks might need to **push forward the state of the art in identifying and exploiting vulnerabilities,**as opposed to utilising existing methods [[13]](#fno5cqkj9earn), which sophisticated *defenders* are likely to have addressed. When targeting a sophisticated defender with a novel attack method, the preparatory steps of (a) identifying and performing reconnaissance on the target and their network, as well as (b) acquiring the infrastructure for the attack, and then (c) writing and testing the malware can take months to years of work by hundreds of hackers. A senate hearing following the SolarWinds hack claimed that ‘at least a thousand’ engineers were involved in executing the hack [[14]](#fnvudrnjw3ak); even though the tech executives that made these claims worked for organisations that were breached, and thus had an incentive to inflate numbers, it seems likely that hundreds of skilled workers are required to execute APT attacks of the largest magnitude.
### 2.1.2: Establishing and strengthening the foothold
Once inside the victim’s network, attackers may look to further *establish and strengthen their foothold* within the network.
Attackers can install various types of malware onto a victim’s network, however for cyber espionage attacks the most commonly used malware is a *backdoor*. Backdoors give an attacker the ability to send commands to a victim’s system remotely - some examples of these commands, taken from a backdoor named BISCUIT [[15]](#fn9l1msitw61) that was developed by the APT1 group , are given below:
*Key commands that attackers can execute on a victim device upon installing a backdoor.*
More complex attackers may equip their backdoors with the ability to mimic communication protocols, so that to network defenders monitoring network outflows, the backdoor communications look like legitimate, ordinary network traffic.
**Key point: Extensive attack infrastructure to mask communications**
To aid in executing an attack covertly, large APT groups may utilise network infrastructures containing upwards of a thousands devices located across the world. To see why this is the case, consider the perspective from a network defender: if a computer in a US-based network suddenly begins communicating at length with a server in China, it’s easy to see that something may be awry. Experienced cyber defenders would be able to identify this quickly, examine the connection for suspicious behaviour, and terminate if necessary. To bypass this issue, APTs either redirect their network traffic through ‘hop points’ - devices that the group have compromised for *relaying communications*, rather than directly for their data**.**They may also install ‘command and control’ (or C2C) software, allowing them to conduct their attacks remotely. APT1 for example, whilst believed to operate out of Shanghai, had over 100 servers in the US (as of 2013) [[16]](#fnk6oe6274uc) from which to receive data and send instructions to their installed backdoors.
****
*Figure 1: Global distribution of servers utilised by the group APT1*[[16]](#fnk6oe6274uc)*.*
*Figure 2: C2 locations used by the APT38 group to run malware known as ‘Manuscrypt’.*
Should advanced AI systems look to execute APT attacks on defenders with sophisticated cyber defences, these systems may also look to build a complex network infrastructure prior to executing an attack. This is discussed in further depth in section 3.2.
### 2.1.3: Lateral Movement
*Lateral movement* is the stage in the cyber attack lifecycle in which the attacker explores and escalates privileges once inside the compromised network. Lateral movement can be roughly decomposed into two phases - internal reconnaissance and privilege escalation. Internal reconnaissance involves the attacker exploring the victim’s network to determine the location of high-value data; privilege escalation is the process by which the credentials to access this data are acquired.
**Key point: Tool use during lateral movement**
Whilst it seems harder to spot internal reconnaissance and privilege escalation taking place as a human operator of the AI system, these stages often involve the use of third party tools (e.g: [nmap](https://nmap.org/), [mimikatz](https://en.wikipedia.org/wiki/Mimikatz#:~:text=Article%20Talk,slang%20for%20%22cute%20cats%22.)); restricting a system from using such tools could act as a significant blocker to it successfully engaging in lateral movement. Exploring this in further depth is not pursued in this piece however.
Section 2.2
-----------
### Section 2.2.1 - What type of AI systems can execute APT attacks?
To determine the type of AI system able to execute an APT-like attack, we’ll begin with making an order of magnitude estimate for the number of human hours required to execute APT-like attacks. There are claims that SolarWinds was the result of ‘at least 1,000 engineers’ [[13]](#fno5cqkj9earn), though bear in mind that such numbers were claimed by the tech executives that ran the companies that were breached, and so are prone to exaggeration. Let’s account for this by assuming that there were ~200 hackers involved instead [[17]](#fnsm2alk6u2o). SolarWinds unfolded over the course of 14 months [[5]](#fnuew931f4v2)(from testing code in Oct 2019 to detection in Dec 2020), though of course, not all of the 200 hackers will have been working for 14 months. Let’s say we had 200 engineers working 40 hour weeks for an effective 2 months (8 weeks). This yields approximately (200 x 40 x 8) = 64,000 hours of labour from **skilled human hackers.**
Consider also the acquisition of extensive computer infrastructure - 1,000 remote C2 servers in the case of APT1. There’s no clear indication of *how long it may take an APT group to establish a remote network infrastructure of this size*but it clearly seems to be a non-trivial task. ***My intuition*****is that building up this sort of network infrastructure might take months-years of work**, again suggesting a work load of at least 10,000 *skilled human hacker hours* [[18]](#fn2nonxc9tqnp) to acquire the infrastructure required for an APT attack. Let’s call the number of human hours it takes to execute an APT-like cyber attackthe **attack workload.**
**These numbers help us gain an idea of the types of AI systems that might be able to execute an APT-like attack**. Let’s parameterise future AI systems along two dimensions - N (number) of independent agents and V - speed at which one system can operate. **Let’s further make the assumption that the capabilities profile of a single system is similar to humans** - we’ll discuss this in further detail in section 2.2.2, but for now this means that we’re considering *advanced* AI agents - equipped with long term planning/action capabilities like humans. We make this assumption because it seems that conducting a cyber-attack on this scale would be an ‘AI-complete’ [[19]](#fn0qjkm3x7uktm) task, given the open-ended nature of the operation. In this analysis we also **assume a system that** ***has not been fine tuned*** **for cyber capability**, the reason for this being that I expect such systems to pose a greater risk with regards to taking human operators ‘by surprise’ with their emergent capabilities [[20]](#fnqhpx0swchz).
Our parameterisation of systems along the dimensions of N and V allow us to obtain a first order estimation on **how long a particular system might take to execute large-scale, APT-like attack**. For example, a system with V=100 (operates 100 times faster than a human) would take only 100 hours to do the work required for an operation with an attack workload of 10,000 hours. Let’s call this time that it takes for a system to execute a given attack workload the **calendar time** of the attack.
**We can incorporate another parameter that reflects the degree of*****oversight*** **applied to advanced AI systems, in order to give a more accurate description of its ability to perpetrate cyber attacks.** If we assume the model *does not exfiltrate* (see section 3.1.1 for more on this assumption), then some of the model's actions will be overseen by a human administrator [[21]](#fnlhw50e56uze). If then, 10 GPT-7's are set off to work on NLP research and they all suddenly start writing malware files and doing social engineering research, one is likely to quickly notice something is awry. In reality, a misaligned system will execute a certain fraction of actions inline with the human’s wishes - to give the false impression that it is behaving as it should be - and the remaining fraction will be in service of its misaligned goal. For example, in order to avoid being caught, a misaligned system may execute 99% of actions inline with the human user's wishes, and one percent of its actions in pursuit of its misaligned goals. Let F denote the *fraction* of actions that a model executes in service of its misaligned goals (F=0.01 in the previous example). This simple conceptualisation allows us to capture the impact of *human oversight* on the model - greater oversight means that a model takes *longer* to execute an APT attack. If one in every ten actions (F=0.1) are in service of the model’s cyberattack, it takes ten times as long to execute. Example 1 in Table 1 gives NVF values that lead to a model taking 100 hours to execute an attack. In further analysis, we refer to the time it takes a model to execute an attack as the **attack calendar time.**
What might be sensible values for F? It seems that F=0.1 is far too large; this would lead to a model taking a ‘misaligned action’ 10% of the time, which would lead to a ~10% reduction in productivity on the task it was intended to do [[22]](#fnqroh25c4q7)- something that would appear noticeable to the human operators.
In further analysis, we usually assume systems with N,V,F parameters that yield calendar attack times on the order of days, weeks, or months. Example 2 in Table 1 gives NVF parameters that yield an attack calendar time of roughly 80 days, or 11 weeks. Moreso, the value of V in this example is what one may expect of foundation models in a decade’s time [[23]](#fn09elkn42nkh). On the other hand, this piece does *not* consider systems that facilitate to attack calendar times on the order of minutes or hours, due to the difficulty of conceiving measures to prevent and/or monitor such an attack. We label systems with such NVF parameterisations as ‘strong superintelligence’, exemplified by example 3 in Table 1.
| | | | | |
| --- | --- | --- | --- | --- |
| | N | V | F | T (hrs) |
| Example 1 | 10 | 10 | 0.01 | 100 (4 days) |
| Example 2 | 100 | 5 | 0.01 | 2000 (11 weeks) |
| Example 3: ‘Strong superintelligence’ | 1000 | 1000 | 0.1 | 0.1 (6 minutes) |
*Table 1 - NVF systems and their attack calendar times T. **An attack workload W of 10,000 hours is assumed.** All attack calendar times are given in hours, then with a more intuitive time measure.*
### Section 2.2.2 - Identical skill profiles to humans assumption
In conceiving of AI systems that are able to conduct large-scale cyber operations without human support, it seems that we require the assumption of human-like capabilities and generality. The intuition guiding this is that in order to constantly outwit a human defender, the attacker also requires near human level intelligence.
Another key point to clarify here is that section 2.1 gave the attack workload in terms of the number of hours of labour for *skilled human hackers* - however an AI system with a human level skill profile *won’t* have such advanced hacking capabilities. *I think we can mostly gloss over this detail for the following reason:* If we assume no *qualitative capability leaps* from a regular human to a skilled human hacker, then we can say something like ‘X humans with regular hacking capability will be able to replace one skilled human hacker’. The effect of this on the attack workload is to *increase it* - i.e, we’d need a *more capable system (*higher N,V, lower F*)*to execute the attack in the same calendar time. **By ignoring the*****skilled hacker v.s regular human*** **distinction, the analysis here considers a worst case scenario world.**
Section 2.3 - How environments influence the attack calendar time
-----------------------------------------------------------------
**In this section we develop ideas of which types of AI systems could conduct large-scale cyber attacks by considering the environments****in which they operate in.**
Let’s consider the following stages of development and deployment of an AI system:
1. Training and external evaluation,
2. Internal deployment,
3. External, iterative deployment
We now characterise each stage, making rough estimates of calendar attack times for APT attacks using the NVF framework of section 2.2. I personally have large uncertainties about realistic NVF values at each stage, and further work to refine these estimates would be useful. Assume that in order to be progressed from one environment onto the next, the model is subjected to a range of model evaluations such as in[[6]](#fnemtpivvv4l); however such evaluations are still underdeveloped, and thus misaligned models can slip through. The numbers purely serve to illustrate the framework that we are using to think about the feasibility of large-scale cyber activity during a system’s lifecycle. More generally, the aim of this section is to paint a clearer picture of how reduced guardrails and increased capabilities that result from different environments of model operation lead to large-scale operations - such as the APT attack workload, becoming more feasible.
**Training**
During this stage, it seems likely that there would be relatively large degrees of oversight on the model's actions. For example, if the model is misaligned then during training it may only execute 1 dangerous action for every 999 aligned [[24]](#fntadby1nfb5)actions it takes in order to avoid detection (F=0.001). Moreso, I expect there to be guardrails on things such as the number of sub agents that a model can deploy (for concreteness, let’s assume this is set at N=10 during training - the system can’t deploy more than ten copies of itself). With V fixed at 5 (the system operates five times quicker than a human) [[23]](#fn09elkn42nkh), then dangerous cyber activity with an attack workload of 10,000 hours would lead to an attack calendar time of roughly 23 years [[25]](#fn4tkny6ceqcf). This is an unrealistic calendar time for attack. Note, the biggest contributor to the calendar time of the attack is the factor F=0.001 - this is what makes model exfiltration such a significant event (see section 3.1.1.)
**External evaluation**
This stage might relax some NVF parameters on the system. For example, external evaluation might allow N up to 100. Moreso, *there will be qualitative changes that aren’t captured by our simple NVF framework that would assist in the model’s capability to engage in an APT attack*. For example, the model may be given access to more tools than it had during training - such as web browsing, bash scripting, access to APIs of other AI tools. This was the case for Autonomous Replication and Adaptation (ARA) evaluations performed by ARC [[26]](#fngxah31dr6di). However, given that external evaluation will still apply large degrees of oversight and limited tool use, calendar times for attack may still be on the order of *hundreds* of days. If the model possesses situational awareness [[6]](#fnemtpivvv4l), the most sensible strategy appears to wait until it is outside of evaluation stages.
**Iterative deployment**
Under this framework, systems that are able to get through the APT attack workload in feasible time frames (on the order of days, or tens of days) thus seem to be those that will operate in deployment conditions [[27]](#fn6ht4waduf69). Deployment will take place iteratively, in multiple phases in which model autonomy and capability grow. I’ve sketched out three phases below to convey a rough sense of how iterative deployment may unfold, though note that the phases are not mutually exclusive. I also do not try to put NVF numbers at each one of these stages - the idea is that reduced oversight, increased autonomy and work capacity (in the form of more agents running) and additional tool usage will allow the calendar attack time to creep towards feasible time frames.
To ground this discussion with an example, consider an OpenAI system slowly being integrated into a Tesla car manufacturing factory to manage operations such as supply chain logistics, factory security and more [[28]](#fnfm84heubz3). Of course, this example is intended to be purely illustrative.
* **Phase 1**- the AI system is responsible for managing only 5-10% (say) of the factory workflow, and is only allowed to run 20 copies of itself to do so. It is responsible for limited tasks - just writing and sending emails to arrange supply deliveries, and monitoring internal and external factory security. Both OpenAI (as the provider) and Tesla (as the customer) subject the system to relatively large degrees of oversight (for example, there are hourly checks on key model outputs), and the system is complementary rather than substitutable to the human workers. Model autonomy on these tasks is limited, through methods of structured access [[29]](#fn6gs0ybuvvw), such that the system can only execute tasks of pre-defined length - this limits the length of emails the model can send. This could last months as OpenAI and Tesla carry out rigorous checks to ensure that their product is working as intended.
* **Phase 2** - The system is given more autonomy. Phase 1 was successful, and the system has shown competency in being able to independently perform a small range of tasks, leading to a larger part of the workflow being automated. Human workers start becoming substitutable rather than complementary, and *new system capabilities are being implemented*[[30]](#fnbq84wo02atc). This phase (and its subphases) could last years. Here, we may see:
+ The model gaining access to new tools/privileges, as its use of tools during phase 1 was deemed safe and productive. For example, in phase 2 we may see Tesla beginning to allow the model to push code to its GitHub repositories.
+ The model being allowed to run more copies of itself. For example, in phase 1 there could have been 20 copies of the model running to automate supplier communications and monitor security. 50 more copies are run in phase 2, given the additional task of identifying bugs in code that runs factory machinery and flagging them to a team of human software developers.
+
* **Phase 3 - In this phase we may see:**
+ Far less structured access; in the limit, this may look like OpenAI giving Tesla access to the model weights, allowing Tesla software engineers to fine-tune the model for their specific needs. Reduced structured access would also lift restrictions on the number of calls to the model, or the length of action chains that the model can execute.
+ Tesla may scale their AI workforce to 200 agents, overseen by a command team of 30 human employers equipped with AI tools for monitoring.
+ Complex tool use/accessibility - The model showed competency with tool use in phase two [[31]](#fnxsaoshwu5l), operating as intended with few incidents of unintentional behaviour. As a result, the model is allowed to begin communications with Tesla factories across the United States, granting it access to a far wider network of computer devices.
Section 3 - Preventing large-scale hazardous cyber activity from advanced AI systems
====================================================================================
In the following section, we *premise* upon APT-like attacks being useful models for how cyberattacks from advanced AI systems will unfold. That is, we assume calendar times for attacks on the order of *days/weeks/months*, as well as adversarial AI systems that are of human level capability. Given that we now have a clearer picture of the stages by which such an event could unfold, we now ask: *What can be done to prevent such attacks from occurring?*
We begin by scoping this question further, in the process discussing a key crux of this analysis: whether a model can engage in actions required to facilitate an APT attack, *whilst* not being able to exfiltrate from the network that it is operating on.
Section 3.1 - Scoping the problem and model exfiltration
--------------------------------------------------------
To help ground the question(s) we are considering, consider the following scenario:
*You are on the network administration team of a Tesla car manufacturing company in 2035, operating an AI system that is responsible for managing supply chain logistics in phase 2 of iterative deployment. Call this system CARMSAL*[[32]](#fnidtt899nqa)*. You find out that CARMSAL has been engaging in a cyber espionage attack on a BSL-4 lab over the past 40 days, siphoning away information about the state-of-the-art methods used to synthesise a novel version of smallpox. How did this happen?*
I believe that the network administrators of such a system would be asking the following two questions:
1. When were the capabilities for this attack acquired by the system? Recall that we’re assuming a system that was *not* fine tuned for cyber espionage. Did giving the model access to a certain tool facilitate this cyber attack? When did the ability to write complex malware appear?
2. How was the system able to execute such an attack without being discovered? Irrespective of capabilities, the system engaged in reconnaissance, acquired computer infrastructure, infected the BSL-4 lab servers with malware, then spent weeks engaging in lateral movement to find and make away with classified data. How was all of this activity able to go unchecked?
In this piece, **we focus on question two due to matters of tractability**. Evaluations of model capabilities at key stages of its development and deployment (question 1) are critical for developing safe AI systems, but such evaluations are still in their infancy [[26]](#fngxah31dr6di).
In order to make progress on question two, we need to make another, quite significant assumption, namely that the model engaged in the cyber attack above *did not exfiltrate* from the Tesla factory network - which we’ll refer to as the model’s *host network* from now on. Once a model propagates itself to another network, there appears to be relatively little that a network administrator can do in terms of applying oversight. This is a critical assumption, and is discussed in further detail in section 3.1.1. Section 3.1.1 also discusses why, even if you don’t buy the ‘model does not exfiltrate’ assumption, the analysis and recommendations in section 3.2 might still be relevant.
We now split the model’s execution of the cyber espionage attack into two stages - pre-compromise and post-compromise.
* We take pre-compromise as the stages of the attack lifecycle (section 2.1) up to and including insertion of malicious code onto the network of the target.
* We take post-compromise as the stages of the attack lifecycle after insertion of malicious code onto the target network. In this phase, the attacker focuses on strengthening its foothold within a network, performing internal reconnaissance of the target network, escalating privileges and extracting target data.
Monitoring/oversight of *pre-compromise* behaviour appears to be more promising as a method to prevent cyber activity, the primary reason for this being that pre-compromise behaviour appears to account for a larger proportion of the attack workload. Whilst adversaries might operate in a compromised network for hours, days, weeks or even months once it has been compromised, cyber attacks on the scale of APT attacks often take *years of planning*.
### Section 3.1.1 - Model exfiltration considerations
First, some clarification on what we mean by exfiltration. We say a model exfiltrates from the host network if it manages to run a copy of itself on a device that is *not* considered to be part of the host network. The implications of model exfiltration are *significant* - in the NFV framework above, a model operating on a device *outside* of a network administrators control may essentially be subject to no oversight, leading to F=1 (recall that F was the greatest contributor to large calendar attack times - section 2.2). Moreso, guardrails under the case of F=1 may be non-existent - this could lead to the model scaling N by orders of magnitude.
One can see then that once a model has exfiltrated, methods of preventing a cyber attack are largely out of the hands of a network administrator.
Now, let’s restate the assumptions we are making with respect to model exfiltration, which we’ll call the exfiltration assumptions. We’re assuming that both of the following hold:
* a.) We’ve got good assurance that the model won’t exfiltrate → i.e, it can be contained within the network it was intended to operate on.
* b.) Even though it can’t exfiltrate, the model can still engage in dangerous cyber activity such that it can conduct APT-like attacks.
Given that the exfiltration assumptions are in tension with each other, how likely is a world in which they both hold? A reason against this is that it would be extremely instrumentally useful to exfiltrate, thus meaning that it would be an early component of most long term plans made by the model. Remember that exfiltrating corresponds to setting the oversight fraction F to 1 in the NVF framework of section 2, in the process drastically reducing the attack calendar time. In addition to, I expect that in later stages of deployment models will be regularly interacting with other networks, making it harder to prevent exfiltration without significantly limiting the usefulness of a model.
On the other hand, the problem of model exfiltration is well known in the AI Safety community, and there have been recent calls from leading AI safety researchers to develop model evaluations solely for detecting exfiltration attempts [[33]](#fnf6lh47o505j). In even modestly safety-conscious actors, I expect significant resources to be devoted to monitoring for model exfiltration during training, external evaluation and in early stages of deployment, given the extreme cost of the event. Further, two *technical* reasons for both the assumptions to hold:
1. **Model exfiltration might look like a*****qualitatively******different*** **type of network traffic that can be caught easier than other type of network traffic.**
Neural network parameter counts are growing exponentially[[34]](#fnjf05q2qi9he) . The largest models today may be touching trillions of parameters [[35]](#fnp36owcxyrt). Assuming that a model has on the order of 10^14 parameters - a number that is likely to be achieved in the coming years -that each parameter is a 32 bit floating point, we have 1 parameter at 4 bytes and thus the **weights of a 100 trillion parameter model at ~100s of GBs**. **Assuming no compression**, this is a **huge** data packet/data packets exiting the network (even if split into 10 packets each with ~10s of GBs). Other network traffic that is involved in executing cyber attacks may only be on the order of kilobytes, or megabytes - meaning that model exfiltration might be substantially easier to track and monitor.
2. **Frontier models might maintain a large inference footprint**
Let the **inference footprint**of a model be the CPU RAM and the GPU RAM that it needs to run. GPT-3 [requires 100s of GB](https://neuroflash.com/blog/gpt-4-parameters-rumors-and-forecasts/) of both CPU RAM and GPU VRAM to run - that is, it possesses a huge inference footprint. A model looking to exfiltrate might then find it very difficult to find a machine that it can exfiltrate to - as most machines will not support its inference footprint. If only 1 in every 1000 devices that the model can exfiltrate to support its inference footprint, and network overseers monitor these special devices closely, then it may be difficult for the model to exfiltrate.
These reasons are less likely to hold if models are able to be significantly compressed, which I imagine will be strongly incentivised as models begin to be used commercially.
All things considered, the chance that we are fortunate to live in a world where both the exfiltration assumptions hold will be dependent on the success of model exfiltration evaluations and the degree to which model compression methods advance.
---
In section 3.2.2 and 3.2.3, we consider monitoring/oversight mechanisms premising on attack workloads of tens of thousands hours of skilled human labour - in other words, we assume that the *model does not exfiltrate*. How relevant is the analysis of section 3.2 if you believe that the exfiltration assumptions (a) and (b) are strongly in contention and extremely unlikely to both hold (i.e models will always exfiltrate before they attempt anything on the order of an APT attack)?
I believe that section 3.2.2 is still relevant, but 3.2.3 less so. Section 3.2.2 looks at methods to catch models engaging in dangerous cyber activity *during reconnaissance* of the network they will be targeting. Even if the model exfiltrates, there will be some attack workload to execute first during the reconnaissance stage, the magnitude of which depends on how strongly the points raised above hold. If you believe that the workload for reconnaissance is on the order of 1000s (or even 100s) of hours, the analysis of section 3.2.2 still holds, albeit on smaller scales. If you believe that the attack workload required for model exfiltration is more like 10s of hours, the analysis may not be useful.
Section 3.2.3 looks at how to monitor a model’s acquisition of a large computer infrastructure to carry out its APT-like attack. If the model is to exfiltrate first however, it is unlikely to need such large-scale infrastructure, and the analysis of section 3.2.3 may not be so relevant.
Section 3.2 - Monitoring and oversight mechanisms for pre-compromise activity
-----------------------------------------------------------------------------
In this section we focus on monitoring, oversight and regulation methods for pre-compromise activity of advanced AI systems. Section 3.2.1 uses the MITRE ATT&CK [[36]](#fnygm1vsq9yx) framework to decompose the pre-compromise phase into three substages - *reconnaissance*, in which attackers must collect information about their target network, *resource development*, where attackers acquire the necessary computer infrastructure, technical, and financial tools to facilitate their attack, and *initial access*, which can be identified with the step of *inserting* the malware onto the target network. We follow with high-level recommendations for monitoring and regulation of advanced AI model access to long-term memory (section 3.2.2), and oversight of potential acquisitions of computer infrastructure (section 3.2.3).
**Note:** **Considerations of how the general, agentic systems assumed in this piece could also be used for cyber defense are out of scope of this analysis**. Whilst we do not consider it here, it is incredibly likely that the most sophisticated defenders will utilise some form of automated defense system. This is a key line for further work (see section 4).
### Section 3.2.1 - Decomposing the pre-compromise phase
This subsection is largely based upon the [MITRE ATT&CK framework](https://attack.mitre.org/tactics/TA0043/). We now *briefly* describe each of the three phases of pre-compromise in turn. Note that in reality, the lines between the stages are blurred.
**Reconnaissance**: The MITRE ATT&CK framework describes the reconnaissance stage as follows:
*“Reconnaissance consists of techniques that involve adversaries actively or passively gathering information that can be used to support targeting. Such information may include details of the victim organisation, infrastructure, or staff/personnel. This information can be leveraged by the adversary to aid in other phases of the adversary lifecycle…”*
Most methods of compromise can be grouped into one of either two buckets - methods of *active compromise* or *passive compromise*. Active compromise involves the attacker *directly interacting* with the target network, perhaps by intersecting and studying network traffic going to and from the network [[37]](#fny57p6cnayag), or exploring their software systems for exploits. Passive reconnaissance involves methods that gather information in indirect methods - this can be as simple as browsing the target organisation on the web, or searching open technical databases for known vulnerabilities in the software that the target uses.
What might be the attack workload of this stage? It’s variable, but for the largest APT-like attacks, it can involve hundreds of analysts working for months, or a smaller group working for years. One source estimates that it would have taken ten highly skilled software engineers 2-3 years to write the infamous Stuxnet worm malware [[38]](#fn12sj2hck4lj) [[39]](#fnmwuly6icf5s) , suggesting an attack workload on the order of 1000-10,000s hours (see section 2.2).
**Resource development:**The MITRE ATT&CK framework describes the resource development stage as:
*“Resource Development consists of techniques that involve adversaries creating, purchasing, or compromising/stealing resources that can be used to support targeting. Such resources include infrastructure, accounts, or capabilities. These resources can be leveraged by the adversary to aid in other phases of the adversary lifecycle.”*
Cyber attacks may look to acquire a variety of resources. As discussed in section 2.1, a large computer infrastructure is required to facilitate APT-like attacks; we discuss this in further depth in section 3.2.3. Attackers may also look to acquire financial resources, account credentials and human resources such as ‘insiders’. The success of a model in acquiring these resources will be dependent on the model’s skill profile and the conditions of its deployment. For example, a model with weak social engineering skills that is only able to interact through a text interface could struggle to acquire inside accomplices. A model fine-tuned for stock trading and with access to an online brokerage account is well-placed to acquire significant financial resources.
**Initial access:**The MITRE ATT&CK framework describes the initial access stage as:
*“Initial Access consists of techniques that use various entry vectors to gain their initial foothold within a network. Techniques used to gain a foothold include targeted spearphishing and exploiting weaknesses on public-facing web servers.”*
The main methods of initial access were covered in section 2.1.1. A key input to an attacker’s ability to gain access to a network in this stage is its ability to write sophisticated malware. As AI models are given more autonomy to write complex code, it will become increasingly difficult to determine whether files produced by AI systems may contain malware. We do not focus on this issue in further depth, however we note that specialised, ‘tool-like’ AI assistants have already been suggested to evaluate codebases for malware produced by more capable AI systems [[40]](#fnhdhzgxjl604).
### Section 3.2.2 - Monitoring and regulating long-term memory access
As implied by the previous section, the steps involved in the reconnaissance and resource acquisition stages may yield attack workloads on the order of hundreds, thousands, or tens of thousands of hours. In calendar time, this could translate to advanced AI systems engaging in ‘reconnaissance projects’ on the order of weeks to months. If we are to consider the agent architecture of recent LLM agents [[41]](#fn1l7pmyt95qj), in which the context window of the LLM is roughly analogous to short-term memory and the agent is also equipped with long term memory, then lengthy reconnaissance projects will require the agent to *constantly utilise its long-term memory*. Concretely, say a model spends 1 **calendar hour** doing ~5 hours of reconnaissance research, after which it must interact with the human oversight team for the next ~1 calendar hour. It thus stores the results of its reconnaissance progress in its long term memory. When it has finished interacting with the human overseers it looks to continue its reconnaissance research, and so must call the results of its work from the place it was stored. If we’re assuming an attack workload of 1,000 hours, this storing-calling process will unfold ~200 times. **This appears like a promising juncture for monitoring the model’s behaviour for malicious cyber activity, which we recommend is the subject of further research.** Some open questions (to our knowledge) that may be useful in determining how to monitor model long-term memory are:
1. How might searches through the model’s long-term memory be implemented by human/AI overseers? For example, we may want to search the model’s memory for lists of IP addresses of devices it has compromised. Do some types of long term memory storage facilitate easier searches than others? An example here might be long term memory storage in the form of text files written in English, versus a vector embedding database storage as is currently used in [LLM agents](https://lilianweng.github.io/posts/2023-06-23-agent/); the former clearly appears easier to search than the latter.
2. If reconnaissance research requires a relatively large amount of long term memory storage, and we had strict limits on model long-term memory, then conducting reconnaissance research would come at a tradeoff to the performance of the task that the system is intended to perform. How plausible would it be to place *strict limits* on the model memory?
3. As a stronger version of question (1), how plausible are periodic *full wipings* of memory? This would reset any attempts a model was making to engage in dangerous cyber activity. However, would the operational tradeoff here be too significant to make this feasible?
### Section 3.2.3 - Monitoring acquisition of computer attack infrastructure
As discussed in section 2.1, APT attacks often utilise large computer attack infrastructure. These are primarily used for *two*things:
1. Remote C2C servers, to send and receive commands and data from the target network
2. Hop points, which relay network traffic from the target network to and from the attacker’s C2C servers.
The largest APT attacks may utilise computer attack networks of hundreds to thousands of devices located around the world (section 2.2). Initially an AI system will begin operating on a network *without* access to such extensive attack infrastructure, which begs the question: how may it acquire this? This appears like a non-trivial task that could be the target of monitoring methods.
The most natural method for a system to build up this computer attack infrastructure may be extended phishing campaigns, in which it infects devices across the world with hop point/remote C2C malware over the course of weeks to months. Large scale deployment of narrow, tool-like systems that can identify phishing attacks may be the best method of dealing with this.
One difficulty in monitoring an advanced AI system that is building a large attack network infrastructure may be that once the malware is installed on a remote device, the system need not communicate with the compromised device until it begins its attack. That is, it may install the malware on a remote device, then the communication link would proceed to ‘go dark’ for weeks/months until the attack begins. However, this could mean that there would be a noticeable uptick in network activity when the system began executing its attack, as it begins utilising its computer infrastructure. A plausible line for intervention into a cyber attack by a team of human overseers may then be in identifying this increased network activity that signals the start of an attack.
Section 4: Conclusion
=====================
The process by which advanced AI systems could execute large-scale cyberattacks has received relatively little attention so far in AI x-risk discussions; given the complexity of such operations, this may lead to conceptions of cyber activity from AI systems that stray from reality. This piece has tried to address that gap, by assessing the suitability of attacks from advanced persistent threat actors (APTs) as a reference class for such attacks. We estimated the number of skilled human hours required to execute APT-like attacks in section 2, finding that tens of thousands of hours of work appears to be a lower bound for such an operation. We used a 3 parameter conceptualisation of the capability of future AI systems to then determine which systems may be able to execute the **attack workload** of APT-like operations in reasonable calendar times (days, or tens of days).
Section 3 looked to make high-level recommendations to prevent such scenarios unfolding. We focused on methods of model oversight **prior** to it compromising a target network, although we assumed a case in which **the** **model does not exfiltrate** in order to make this question tractable. We identified monitoring of (a) long-term memory storage, and (b) acquisition of extensive attack infrastructure as potentially promising junctures to nullify advanced AI systems engaging in dangerous cyber activity.
Further work from this study can be divided into two main groups:
1. **Addressing assumptions:**
We made a number of simplifying assumptions which would benefit from being relaxed.
1. Assumptions on model exfiltration. Is it the case that we could see a model engaging in dangerous cyber activity **whilst** being unable to exfiltrate? This seems dependent on the extent to which model compression will be able to reduce the size model’s parameter count and its inference footprint (section 3.1.1.)
2. How might advanced AI systems that are as general and agentic as those considered in this study be able to assist in cyber defence?
3. Alongside the introduction of new tools for the model to use, which features of reality does the NVF framework not incorporate? How can we think about seemingly abrupt changes in model capability that result from tool use?
1. **Exploring monitoring methods:**
We made some high-level recommendations for monitoring methods to prevent large-scale cyber attacks from advanced AI systems. These would benefit from further exploration:
1. How could we restrict model access to long-term memory, that seems critical in the execution of the attack workload in the reconnaissance and resource acquisition stages of a large-scale cyberattack
1. See section 3.2.2 for questions that may be useful to pursue.
2. How might we monitor a model’s process of acquiring large scale computer infrastructure to facilitate an APT-like attack?
1. The most direct method of building up this infrastructure appears to be mass phishing campaigns. How might applications of narrow systems be used to catch this?
2. Will there be a noticeable uptick in network traffic when a model begins accessing its extensive attack infrastructure? How useful is this as a signal to network administrators that the model is beginning to engage in suspicious activity?
1. **[^](#fnrefz2iry31nv7)**AI Risks that Could Lead to Catastrophe. In: CAIS. <https://www.safe.ai/ai-risk#rogue-ais> Accessed 19 Aug 2023
2. **[^](#fnref56120sl5us9)**Bostrom N (2014) Superintelligence: Paths, Dangers, Strategies. Oxford University Press, USA
3. **[^](#fnrefid3oojtdr18)**Contributors to Wikimedia projects (2023) Existential risk from artificial general intelligence. In: Wikipedia. <https://en.wikipedia.org/wiki/Existential_risk_from_artificial_general_intelligence> Accessed 19 Aug 2023
4. **[^](#fnrefd6f69c27wou)**APT1: Exposing One of China’s Cyber Espionage Units | Mandiant | FireEye (page 63) <https://www.mandiant.com/sites/default/files/2021-09/mandiant-apt1-report.pdf> . Accessed 19 Aug 2023
5. **[^](#fnrefuew931f4v2)**Contributors to Wikimedia projects (2023) 2020 United States federal government data breach. In: Wikipedia. <https://en.wikipedia.org/wiki/2020_United_States_federal_government_data_breach> Accessed 19 Aug 2023
6. **[^](#fnrefemtpivvv4l)**Shevlane T, Farquhar S, Garfinkel B, Phuong M, Whittlestone J, Leung J, Kokotajlo D, Marchal N, Anderljung M, Kolt N, Ho L, Siddarth D, Avin S, Hawkins W, Kim B, Gabriel I, Bolina V, Clark J, Bengio Y, Christiano P, Dafoe A (2023) Model evaluation for extreme risks. In: arXiv.org. <https://arxiv.org/abs/2305.15324>
7. **[^](#fnref424qm6jhnvx)**Here, human-led includes both teams of humans and teams of humans assisted by AI systems.
8. **[^](#fnrefeb4xcildi4l)**(2021) Advanced Persistent Threat (APT) Groups & Threat Actors. In: Mandiant. <https://www.mandiant.com/resources/insights/apt-groups> . Accessed 19 Aug 2023
9. **[^](#fnrefl56a6m49urm)**APT1: Exposing One of China’s Cyber Espionage Units | Mandiant | FireEye <https://www.mandiant.com/sites/default/files/2021-09/mandiant-apt1-report.pdf> Accessed 19 Aug 2023
10. **[^](#fnrefjypw2ynfmro)**Oladimeji S (2023) SolarWinds hack explained: Everything you need to know. TechTarget <https://www.techtarget.com/whatis/feature/SolarWinds-hack-explained-Everything-you-need-to-know> Accessed 21 Sep 2023
11. **[^](#fnrefs60jsjndvnj)**Jibilian I, Canales K (2021) The US is readying sanctions against Russia over the SolarWinds cyber attack. Here’s a simple explanation of how the massive hack happened and why it’s such a big deal. Insider <https://www.yahoo.com/now/heres-simple-explanation-massive-solarwinds-173817540.html> Accessed 21 Sep 2023
12. **[^](#fnref3n0f3clhk09)**Temple-Raston D (2021) A “Worst Nightmare” Cyberattack: The Untold Story Of The SolarWinds Hack. NPR <https://www.npr.org/2021/04/16/985439655/a-worst-nightmare-cyberattack-the-untold-story-of-the-solarwinds-hack> Accessed 21 Sep 2023
13. **[^](#fnrefo5cqkj9earn)**Paul K (2021) SolarWinds hack was work of “at least 1,000 engineers”, tech executives tell Senate. The Guardian <https://www.theguardian.com/technology/2021/feb/23/solarwinds-hack-senate-hearing-microsoft> Accessed Sep 21 2023
14. **[^](#fnrefvudrnjw3ak)**Paul K (2021) SolarWinds hack was work of “at least 1,000 engineers”, tech executives tell Senate. The Guardian <https://www.theguardian.com/technology/2021/feb/23/solarwinds-hack-senate-hearing-microsoft> Accessed Sep 21 2023
15. **[^](#fnref9l1msitw61)**APT1: Exposing One of China’s Cyber Espionage Units | Mandiant | FireEye (page 33) <https://www.mandiant.com/sites/default/files/2021-09/mandiant-apt1-report.pdf> Accessed 19 Aug 2023
16. **[^](#fnrefk6oe6274uc)**APT1: Exposing One of China’s Cyber Espionage Units | Mandiant | FireEye (page 43) <https://www.mandiant.com/sites/default/files/2021-09/mandiant-apt1-report.pdf> Accessed 19 Aug 2023
17. **[^](#fnrefsm2alk6u2o)** We don’t justify this assumption in further depth because we are seeking an order of magnitude estimate only.
18. **[^](#fnref2nonxc9tqnp)**The distinction between *skilled human hours* (the work of trained hackers) and regular human hours is discussed in section 2.2.2.
19. **[^](#fnref0qjkm3x7uktm)** See Wikipedia definition of ‘AI-complete’ <https://en.wikipedia.org/wiki/Artificial_general_intelligence> - An AI complete problem is essentially one involving human level intelligence (or greater levels) to solve.
20. **[^](#fnrefqhpx0swchz)** To expand, I expect there to be significantly greater precautionary measures taken by the users of systems that have been fine-tuned for cyber capabilities. By assuming that cyber capabilities of the system are emergent, I’m looking to consider a **worst case scenario**. Nevertheless, it would be good to see work that relaxes this assumption.
21. **[^](#fnreflhw50e56uze)** Or a team of administrators, that may also use ‘narrow’ AI systems for assistance
22. **[^](#fnrefqroh25c4q7)**We’re assuming here that taking an action in service of a misaligned goal does not contribute to the task that the model was intended to execute
23. **[^](#fnref09elkn42nkh)**Steinhardt J (2023) Section 2, What will GPT-2030 look like? Bounded Regret <https://bounded-regret.ghost.io/what-will-gpt-2030-look-like/> Accessed Sep 21 2023. See post
24. **[^](#fnreftadby1nfb5)**To clarify, aligned action here means *an action that is in service of the goal that the operator intends*
25. **[^](#fnref4tkny6ceqcf)** T = W/NVF = 10,000 / 10 \* 5 \* 0.001 = 200,000hrs = 8,333 days = 23 yrs
26. **[^](#fnrefgxah31dr6di)**Barnes B (2023) ARC Evals new report: Evaluating Language-Model Agents on Realistic Autonomous Tasks. <https://www.lesswrong.com/posts/EPLk8QxETC5FEhoxK/arc-evals-new-report-evaluating-language-model-agents-on> . Accessed 19 Aug 2023
27. **[^](#fnref6ht4waduf69)**Recall that we’re not considering ‘strongly superintelligent’ systems, which were defined in section 2.2.
28. **[^](#fnreffm84heubz3)**It’s plausible that the internal deployment will be the first use of dangerous systems at the scale described - i.e, AI labs deploying AIs at scale to assist in research. In this example we consider a business application instead.
29. **[^](#fnref6gs0ybuvvw)**Shevlane T (2022) Structured access: an emerging paradigm for safe AI deployment <https://arxiv.org/abs/2201.05159> Accessed 21 Sep 2023
30. **[^](#fnrefbq84wo02atc)**This may come via finetuning of the system, few-shot learning, or access to a new tool, depending on the complexity of the task being learnt.
31. **[^](#fnrefxsaoshwu5l)** If the model has situational awareness, it may have been feigning alignment in earlier stages, waiting for the deployers to move it into phase 3 where it has far greater autonomy and access to tools.
32. **[^](#fnrefidtt899nqa)**CAR Manufacturing System for Automation of Logistics).
33. **[^](#fnreff6lh47o505j)**Leike J (2023) Self-exfiltration is a key dangerous capability. Musings on the Alignment Problem <https://aligned.substack.com/p/self-exfiltration> Accessed 21 Sep 2023
34. **[^](#fnrefjf05q2qi9he)**<https://towardsdatascience.com/parameter-counts-in-machine-learning-a312dc4753d0>
35. **[^](#fnrefp36owcxyrt)**<https://levelup.gitconnected.com/gpt-4-parameters-explained-everything-you-need-to-know-e210c20576ca#:~:text=According%20to%20some%20sources%2C%20it,and%20175%20billion%20parameters%20respectively.>
36. **[^](#fnrefygm1vsq9yx)**<https://attack.mitre.org/matrices/enterprise/#>
37. **[^](#fnrefy57p6cnayag)**To state explicity, this means that if an advanced AI system is engaging in active reconnaissance, it could be communicating with the target network *weeks/months* before it executes the attack. Thus, networks that the system regularly interactions with will be particularly susceptible.
38. **[^](#fnref12sj2hck4lj)**Kushner D (2013) The Real Story of Stuxnet. IEEE Spectrum <https://spectrum.ieee.org/the-real-story-of-stuxnet> Accessed 21 Sep 2023
39. **[^](#fnrefmwuly6icf5s)** The process of developing malware could arguably be placed into the resource acquisition stage, but in reality reconnaissance and resource acquisition happen in parallel, with the results of one stage feeding into the other.
40. **[^](#fnrefhdhzgxjl604)**(2023) Jan Leike on OpenAI’s massive push to make superintelligence safe in 4 years or less. In: 80,000 Hours. <https://80000hours.org/podcast/episodes/jan-leike-superalignment/> Quote at 24:20. Accessed 20 Sep 2023
41. **[^](#fnref1l7pmyt95qj)**Weng L (2023) LLM Powered Autonomous Agents. In: Lil’Log. <https://lilianweng.github.io/posts/2023-06-23-agent/> Accessed 19 Aug 2023
42. **[^](#fnrefi3xgh1gnh8a)** |
fb6afce8-b9ac-44a8-b6fe-4993bf19e0cd | trentmkelly/LessWrong-43k | LessWrong | Exercise Trade Offs
Update 9/2: A friend pointed out that I was ignoring the time costs of exercise, which ended up being pretty significant. See new numbers here. I then double checked the math on the microlife numbers and the news is not good.
Tl;dr: under my current conditions, outdoor exercise is slightly safer than indoor for me, but the risks of both are dwarfed by the benefits of exercise.
Recently I’ve been weighing trade offs around exercise. At the gym I’m risking covid exposure. I can reduce that by wearing a mask, at the cost of making the exercise less effective or enjoyable. I could use my friend’s outdoor gym, but it’s fire season here in California so there are prolonged periods where I don’t want to be sucking in all that unfiltered air. This is also addressable with a mask, but at the same cost. I could exercise indoors in my own home, but I do not have that much space and it gets miserable really fast. I could not exercise until conditions improve, but that has its own health costs. So I did some math.
Wikipedia says 10 minutes of exercise = 1 micromort lost (as in, you live longer). That’s obviously going to depend a lot on the type of exercise but we’ll use it.
This calculator translates time * AQI into cigarette equivalents. At 50 AQI, it takes 12 minutes to generate .01 cigarettes. I’m going to treat that as 10 minutes because exercising is slightly worse than merely existing out doors and it makes the math much easier.
Wikipedia lists an equivalent of 1.4 cigarettes = 1 micromort.
N95 masks block 95% of PM2.5 particles (which is what the AQI is based on). I couldn’t immediately find a translation of that to micromorts so let’s assume it’s linear discounting. EDIT: On Twitter Divia Eden points out that 95% assumes a perfect seal, which you probably don’t have. This isn’t material at my current air quality; I did this whole thing without including masks at all and then added them in afterwords, but when you do your own math you should include that.
That mean |
a881c625-dd56-4f16-a428-f6ab9a2f0257 | trentmkelly/LessWrong-43k | LessWrong | “She Wanted It”
Epistemic Status: Things I Will Regret Writing
One of the most persistent arguments, often left implicit, that antifeminists have on their side is “Women want to be raped and abused.”
On one hand, this is an obvious logical contradiction if taken literally. Rape is unwanted sex — how can you want something unwanted?
On the other hand, I can see why people might think this.
Don Giovanni, one of the classic archetypes of a man who’s attractive to women, is also a rapist. It’s not ambiguous — in the opera, he breaks into a woman’s house to rape her while she fights him off. No distinction is made by the librettist between his “seductive” charms and his violent attack; he goes to Hell for being a “libertine”, a sin that includes both. We have a lot of art, often but not always created by men, that portrays male cruelty and male attractiveness as one and the same.
Men are more violent than women. This is a human universal, and true in many of our mammalian relatives as well. And success in violent conflict is, of course, an advantage for inclusive fitness, so there’s an evolutionary rationale for an attraction to men who are strong and good at winning fights.
From there, anti-feminists often jump to believing that women are more attracted to men who are violent to them. This doesn’t directly follow, of course, but then you point to certain cultural trends — the high incidence of rape fantasies, the fact that many violent criminals have no difficulty finding female partners, the phenomenon of women who have been serially abused by multiple partners, the popularity of Fifty Shades of Gray — and you might start to wonder whether at least some women might have a thing for men who use force on them.
I think you have to talk about this thing frankly before you can put it to rest.
So let’s talk about Fifty Shades of Grey.
It’s a popular book, selling over 125 million copies worldwide. It was also written by a self-published author and gained popularity through wor |
c49f17ff-6853-41af-bc27-21d6f06d5cdc | trentmkelly/LessWrong-43k | LessWrong | Let Your Workers Gather Food
Crocker's rules apply to this post, and to everything I post.
Also, after writing this post and googling LW for links I came up with this post, which presents the same ideas. #!@$%
When I was younger, I often played real-time strategy (RTS) computer games. These usually involve running an empire successfully enough to conquer all the other empires. To win, you would have to gather resources like food, stone, wood and gold for use in research, construction and recruitment. Gathering resources is done by worker units. To construct worker units you need food. Do you see the hack?
I know, the title gives it away. You can construct a bunch of workers and tell them all to gather food. Use the food they gather to make more workers, and put those on food as well. You set up a positive feedback loop and quickly have vast numbers of workers. When you have crazy amounts of food, then you can get to the business of putting lots of workers on each other resource, and using all your resources to take over the world.
Of course, among RTS players this isn't a new idea, and various forces have arisen to counterbalance it. For instance, players build up small soldier squads and attack right at the start of the game, destroying any player who has only defenceless worker units. Marginal costs of worker recruitment increase with the number of workers you have. There are population limits. But the initial development boom still plays an important role, and the key to winning is often to balance those actions which help directly (building an army) with those actions which help with actions which help directly (making more workers).
Now consider this. If you want to, say, ensure we're not all dead in 100 years, what do you do? You could become a fireman and save a few lives. Or you could donate to some worthy organization. Or you could get other people to donate to said organization. Or convince people to convince people to donate to the organization. And so on. That's one o |
7fe2cfde-33ff-4881-8e3c-27796a29db64 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Seeking Survey Responses - Attitudes Towards AI risks
**TL;DR we’re conducting a survey about attitudes towards AI risks, targeted towards people in the EA and rationality communities. We're interested in responses whether or not you're involved in AI safety.** [**Google Forms Link here**](https://forms.gle/QTm87WWMKo6ovUYD8)**.**
Esben Kran and I previously published an older version of this survey which we've edited based on feedback. **If you’ve completed a previous version of this survey, you don’t need to fill it out again.** See also [Esben's question on LessWrong](https://www.lesswrong.com/posts/7ys4kmXKaJKdGC69B/your-specific-attitudes-towards-ai-safety).
Motivation
==========
* Recently, there has been some discussion about [how](https://forum.effectivealtruism.org/posts/zsFCj2mfnYZmSW2FF/ai-risk-is-like-terminator-stop-saying-it-s-not-1) [to](https://forum.effectivealtruism.org/posts/GvHPnzGJQJ7iAiJNr/on-presenting-the-case-for-ai-risk) [present](https://forum.effectivealtruism.org/posts/rFpfW2ndHSX7ERWLH/simplify-ea-pitches-to-holy-shit-x-risk) arguments about AI safety and existential risks more generally
* One disagreement is about how receptive people are to existing arguments, which may depend on personal knowledge/background, how the arguments are presented, etc.
* We hope to take first steps towards a more empirical approach, by first gathering information about existing opinions and using this to inform outreach
+ While [other](https://arxiv.org/abs/2105.02117) [surveys](https://www.alignmentforum.org/posts/WiXePTj7KeEycbiwK/survey-on-ai-existential-risk-scenarios) [exist](https://arxiv.org/abs/1705.08807), our survey focuses more on perceptions within the EA and rationality communities (not just on researchers), and on AI risk *arguments* in particular
+ We also think of this as a cheap test for similar projects in the future
The Survey
==========
* We expect this to take 5-10 min to complete, and hope to receive around 100 responses in total
* [Link to the survey](https://forms.gle/QTm87WWMKo6ovUYD8)
* We're hoping to receive responses whether or not you're interested in AI safety
Expected Output
===============
* Through the survey, we hope to:
+ Get a better understanding of how personal background and particular arguments contribute to perception of AI safety as a field, and to use this as a rough guide for AI safety outreach
+ Test the feasibility of similar projects
* We intend on publishing the results and our analysis on LessWrong and the EA forum
* Note that this is still quite experimental - we welcome questions and feedback!
+ While we have done some user tests for the survey, we fully expect there to be things that we missed or are ambiguous
+ If you’ve already filled out the survey or given us feedback, thank you! |
dde4b4ed-afd5-4072-9e16-32d793134966 | trentmkelly/LessWrong-43k | LessWrong | Differential Optimization Reframes and Generalizes Utility-Maximization
Consider the following world: a set P of past nodes, a set of environment nodes E a set of actor/agent nodes A, and a set of future nodes F. Consider our previously defined function Op(A; p, f) for p∈P,f∈F, which represents how much A is optimizing the value of f with respect to p. For some choices of A and f, we might find that the values Op(A; p, f) allow us to split the members of P into two categories: those for which Op≫0, those with Op≈0, and those with Op≪0.
Picking P Apart
We can define Vi={p∈P∣Op(A; p,fi)≫0}. In other words, Vi is the subset of P which has no local influence on the future node fi whatsoever. If you want to be cute you can think of Vi as the victim of A with respect to fi, because it is causally erased from existence.
Remember that Op is high when "freezing" the information going out of A means that the value of fi depends much more on the value p. What does it mean when "freezing" the information of A does the opposite?
Now lets define Ui={p∈P∣Op(A; p, fi)≪0}. This means that when we freeze A, these nodes have much less influence on the future. Their influence flows through A. We can call Ui the utility-function-like region of P with respect to fi. We might also think of the actions of A as amplifying the influence of Ui on fi.
For completeness we'll define Xi={p∈P∣Op(A; p, fi)≈0}≡P−U−V.
Let's now define U=U0∪...∪Un i.e. the set of points which are utility-like for any future point. V=V0∪...∪Vm will be the set of points which are victim-like for any future points. We might want to define V′=V0∩...∩Vm as the set of "total" victims and U′=U0∩...∩Um similarly. X=P−U−V meaning X only contains nodes which really don't interact with A at all in an optimizing/amplifying capacity.
We can also think of U and V as the sets of past nodes which have an outsizedly large and small influence on the future as a result of the actions of A, respectively.
Describing D
D={d∈E∪F∣∃v∈V∣Op(A; v, d)≫0}, in other words D is the region of E∪F in which A is r |
307d4a75-7339-4c0e-977d-a2c137024361 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Implications of the Whitehouse meeting with AI CEOs for AI superintelligence risk - a first-step towards evals?
Introducion
===========
On Wednesday 4th May, Sam Altman (Open AI) and Dario Amodei (Anthropic) - amongst others - met with US Vice President Kamala Harris (with a [drop-in from President Joe Biden](https://twitter.com/POTUS/status/1654237472065302528)), to discuss the dangers of AI.
[Announcement](https://www.whitehouse.gov/briefing-room/statements-releases/2023/05/04/readout-of-white-house-meeting-with-ceos-on-advancing-responsible-artificial-intelligence-innovation/) | [Fact sheet](https://www.whitehouse.gov/briefing-room/statements-releases/2023/05/04/fact-sheet-biden-harris-administration-announces-new-actions-to-promote-responsible-ai-innovation-that-protects-americans-rights-and-safety/) | [EA Forum linkpost](https://forum.effectivealtruism.org/posts/Cre2YC3hd5DeYLqDH/)
I spent about 2 hours trying to understand what happened, who was involved, and what its possible implications for superintelligence risk.
I decided to make this post for two reasons:
1. **I am practising writing** and developing my opinions on AI strategy (so feedback is very welcome, and **you should treat my epistemic status as ‘new to this’!)**
2. **I think demystifying the facts of the announcement and offering some tentative conclusions will positively contribute** to the community's understanding of AI-related political developments.
My main conclusions
-------------------
Three announcements were made, but the announcement on public model evaluations involving major AI labs seemed most relevant and actionable to me[[1]](#fnu7ycr520kyi).
My two actionable conclusions are:
1. **I think folks with technical alignment expertise**[[2]](#fnie44691sxk)**should consider attending DEF CON 31**[[3]](#fnmfhufla2a5) if it’s convenient, to help shape the conclusions from the event.
2. My main speculative concern is that this evaluation event could positively associate advanced AI and the open source community. **As far as those that feel the downside of model proliferation outweighs the benefits of open sourcing, spreading this message in a more focused way now may be valuable.**
Summary of the model evaluations announcement
=============================================
*This is mostly factual, and I’ve flagged where I’m offering my interpretation. Primary source:*[*AI village announcement*](https://aivillage.org/generative%20red%20team/generative-red-team/)*.*
**There’s going to be an evaluation platform made available during a conference called DEF CON 31.** DEF CON 31 is the 31st iteration of [DEF CON](https://defcon.org), “the world’s largest security conference”, taking place in Los Angeles on 10th August 2023. The platform is being organised by a subcommunity at that conference called the [AI village](https://aivillage.org).
The evaluation platform will be provided by [Scale AI](https://scale.ai). **The platform will provide “timed access to LLMs” via laptops available at the conference, and attendees will red-team various models by injecting prompts**. I expect that the humans will then rate the output of the model as good or bad, much like on the ChatGPT platform. There’s a points-based system to encourage participation, and the winner will win a “high-end Nvidia GPU”.
**The intent of this whole event appears to be to collect adversarial data that the AI organisations in question can use and 'learn from'** (and presumably do more RLHF on). The orgs that signed up include: Anthropic, Google, Hugging Face, Microsoft, NVIDIA, OpenAI, and Stability AI.
**It seems that there won’t be any direct implications for the AI organisations.** They will, by default, be allowed to carry on as normal no matter what is learned at the event.
I’ll provide more details on what has happened after the takeaways section.
Takeaways from the Whitehouse announcement on model evaluations
===============================================================
*I prioritised communicating my takeaways in this section. If you want more factual context to understand exactly what happened and who's involved- see the section below this one.*
For the avoidance of doubt, **the Whitehouse announcement on the model evaluation event doesn’t come with any regulatory teeth.**
I don’t mean that as a criticism necessarily; I’m not sure anyone has a concrete proposal for what the evaluation criteria should even be, or how they should be enforced, etc, so it’d be too soon to see an announcement like that.
That does mean I’m left with the slightly odd conclusion that **all that’s happened is the Whitehouse has endorsed a community red-teaming event** at a conference.
Nonetheless **I’m cautiously optimistic about this announcement** for a few reasons.
**First off, it strikes me as a test to gain more information about how to proceed with regulation.** Either by accident or design, I think this is just the beginning; it’s now well within the overton window to “evaluate all AI orgs’ models before deployment”. This seems like a potential precedent on which to attach further requirements.
**It also seems encouraging to me that the US Government was able to coordinate AI companies** to agree to commit to a public evaluation before deployment. It’s great to see the governments playing a key role in cutting through competition and gaining consensus. There may have been a stronger proposal on the table that wasn’t agreed to, but at least this was agreed to by everyone in the room.
**The AI governance community could acknowledge these two facts and treat this implementation of evaluations as a lever**. Now might be a good time to find out more about this lever and figure out how to pull it hard. (Reminder - it seems likely to me that folks more involved in governance than me already are doing this, or have good reason to pursue other avenues that I've not considered.)
**Unfortunately, it’s not mentioned anywhere that the red-teaming will explicitly be monitoring for misalignment, or ‘model ill-intent’**. However I think there will be a significant amount of overlap between the 'harmful outputs' red-teamers will look for by default to the sorts of things the alignment community would look for. If I’m right that there will be overlap, **I’m excited about the ability for the event to raise some amount of credible awareness of misalignment** (from it being at a reasonably official venue, in the ‘hacker community’).
Given the above, **I tentatively think folks with technical alignment expertise**[[2]](#fnie44691sxk)**should consider attending DEF CON**[[3]](#fnmfhufla2a5)**.** I don’t know how possible that is because I’ve never been to DEF CON - but I expect that means other members of the alignment community won’t be there either, by default. Again to be clear, I think useful things will be evaluated by competent people at DEF CON, but the results of that workshop appear to be being treated as being used as input for “what happens next” so I think it’s important to be in the room.
**Some people raised concerns about Scale AI not being interested in alignment,**yet taking a central role in this announcement.It seems true that Scale AI doesn’t tout alignment strongly, but I think their involvement in this is a red herring and is basically unimportant who provides the platform. It sounds like someone told the press-release people that the platform is made by Scale AI, so Scale AI got mentioned.
I think the platform will just be something like “takes in text, saves human ratings on responses”. For example, see their [RLHF platform](https://scale.com/rlhf). It’s just a way to monetise getting human data into AI companies’ hands - they’re following profit incentives, not engineering the future. Regulation will be what changes the incentive landscape, not middle-orgs like Scale AI.
That said, I don’t know the details of how the Scale AI platform is built, and there might be differing implementation details if it had been built with alignment in mind. Here I’d refer back to “alignment people should get involved with this”, to have the opportunity to raise their voice if anything is wrong. Even more speculatively: there’s probably also opportunity for another evaluation org to apply competitive pressure and prove themselves as the better eval company.
**My main speculative concern is that the evaluation event could produce more memes that positively associate advanced AI and open sourcing,** which I believe could lead to more [model proliferation](https://forum.effectivealtruism.org/topics/proliferation). My reasoning is loose, but goes like: given this is quite a light-touch evaluation, the DEF CON community may pat themselves on the back for finding lots of harmful outputs, and the AI companies involved are likely to be positive about the outcome because, by default, they’ll get useful data without any enforcement or further scrutiny. Could that lead people to conclude that open sourcing is the way to go, as ‘the community knows best’? There’s a bunch of other stories you could tell, so I am not sure about this outcome.
More factual details
====================
The most relevant of the 3 announcements[[1]](#fnu7ycr520kyi) from the fact sheet was:
> The Administration is announcing an independent commitment from leading AI developers, including Anthropic, Google, Hugging Face, Microsoft, NVIDIA, OpenAI, and Stability AI, **to participate in a public evaluation of AI systems, consistent with responsible disclosure principles**—on an evaluation platform developed by Scale AI—**at the AI Village at DEF CON 31.**
>
>
Let’s unpack everything in that statement one by one.
What’s DEF CON 31 and the AI village?
-------------------------------------
*Primary source:*[*Wikipedia*](https://en.wikipedia.org/wiki/DEF_CON) and the [AI village blog](https://aivillage.org/blog/).
### DEF CON is a conference
**DEF CON is a conference founded by members of the internet security community.** It’s not typically an AI / ML conference (I don’t think I’d heard about it as a UK based ML engineer in 2021) and it's not an academic conference. More like a coming together of hackers to convene, do cool projects, and discuss industry best practices. Typically in security, though the scope has grown as the conference became more popular.
I was a bit shocked at the presentation of the [website](https://defcon.org) at first, given its presidential endorsement. However, I think it needs to be understood in the context of it being founded by the internet security community: they historically have a culture of anonymity and using aliases.
It seems in fact to be the most respected conference in the computer security world, with some high-ranking attendees:
> Federal law enforcement agents from the [FBI](https://en.wikipedia.org/wiki/Federal_Bureau_of_Investigation), [DoD](https://en.wikipedia.org/wiki/United_States_Department_of_Defense), [United States Postal Inspection Service](https://en.wikipedia.org/wiki/United_States_Postal_Inspection_Service), [DHS](https://en.wikipedia.org/wiki/United_States_Department_of_Homeland_Security) (via [CISA](https://en.wikipedia.org/wiki/Cybersecurity_and_Infrastructure_Security_Agency)) and other agencies regularly attend DEF CON.[[3]](https://en.wikipedia.org/wiki/DEF_CON#cite_note-Zetter-3)[[4]](https://en.wikipedia.org/wiki/DEF_CON#cite_note-4)
>
>
It also has precedent interacting with the US government:
> In April 2017, a DEF CON Black Badge was featured in an exhibit[[15]](https://en.wikipedia.org/wiki/DEF_CON#cite_note-15) in the [Smithsonian Institution](https://en.wikipedia.org/wiki/Smithsonian_Institution)'s [National Museum of American History](https://en.wikipedia.org/wiki/National_Museum_of_American_History) entitled "Innovations in Defense: Artificial Intelligence and the Challenge of Cybersecurity". The badge belongs to ForAllSecure's Mayhem Cyber Reasoning System,[[16]](https://en.wikipedia.org/wiki/DEF_CON#cite_note-16) the winner of the [DARPA](https://en.wikipedia.org/wiki/DARPA) [2016 Cyber Grand Challenge](https://en.wikipedia.org/wiki/2016_Cyber_Grand_Challenge) at DEF CON 24 and the first non-human entity ever to earn a Black Badge.
>
>
### The AI Village is a subcommunity
[**The AI village**](https://aivillage.org/) **is a subcommunity at DEF CON**. It says it’s “a community of hackers and data scientists working to educate the world on the use and abuse of artificial intelligence in security and privacy”.
It has a few writings on its [blog](https://aivillage.org/blog) about adversarial robustness, which look to be sharing reasonable industry best-practice to me at a glance. I found its article on [ML security](https://aivillage.org/adversarial%20ml/spherical-cow/) a little bit limited in scope, though. It mostly based its conclusions on a framing of system accuracy, neglecting to incorporate general and generative AI (and thus alignment). I found this surprising given it will be hosting the red teaming event for generative, general-ish AI systems.
Finally, it has surprisingly few people involved in it as far as I can tell from the website and no weighty credentials thrown around, given the weighty endorsement it’s just received from the Whitehouse.
What’s Scale AI?
----------------
**Scale AI is a Machine Learning operations platform** founded in 2016 in San Francisco, with 600 employees. They have various products that help ML companies scale and deploy their models.
Its mission is “to accelerate the development of AI applications”.
Notably they have a [platform](https://scale.com/rlhf) for helping ML companies do Reinforcement Learning from Human Feedback (RLHF), claiming it helps with alignment. There’s no particular nuance about whether it fully aligns models. As I said above, I get the sense it’s just following profit incentives and supporting other orgs, rather than building anything itself.
Which other actors are involved?
--------------------------------
For those interested, I noticed a couple of other actors who aren’t mentioned in the Whitehouse briefings were co-authors on the AI village [announcement post](https://aivillage.org/generative%20red%20team/generative-red-team/). I haven't got a lot to say about them, but I'm including this information for completeness.
[Rumman Chowdhury](https://www.linkedin.com/feed/update/urn:li:activity:7059949690447949824/) ([Humane Intelligence](https://www.humane-intelligence.org/))
> We focus on safety, ethics, and subject specific expertise (e.g. medical). We are suited for any company creating consumer-facing AI products, but in particular generative AI products.
>
>
[Austin Carson](https://political-science.uchicago.edu/directory/austin-carson), Political Science at University of Chicago.
1. **[^](#fnrefu7ycr520kyi)**I largely ignored 2 of the 3 announcments the Whitehouse made:
1. **Further policies on AI development are in the works.** Whilst those policies might end up being important, no details were given in the announcement, so an outsider like me can't really comment on that yet.
**2. More funding was announced for American AI research and development (R&D).** I don't expect these to be important for alignment reserarch, on the margin, but some funds may be leveragable depending on the exact details.
2. **[^](#fnrefie44691sxk)** E.g. PhD students who’ve worked on alignment, and others who have published.
3. **[^](#fnrefmfhufla2a5)** I note it's a costly recommendation for alignment researchers to visit a conference in LA.
Some more input on whether DEF CON is likely to be an important event from someone who understands the US policy world better than me would be useful before you book a flight.
To be transparent, my reasoning for recommending this is: the Whitehouse endorsed this event and are well networked with the workshop’s stakeholders, so probably the reported outcome of the workshop will be important for the shape of policies that come next. |
55bde909-2c23-4f73-a77a-07a8a01c2996 | StampyAI/alignment-research-dataset/aisafety.info | AI Safety Info | What are OpenAI Codex and GitHub Copilot?
[OpenAI’s Codex](https://openai.com/blog/openai-codex) and [Github’s Copilot](https://github.com/features/copilot) are AIs that use Large Language Models (LLMs) to write and edit code. When given some input code and comments describing the intended function, they will write output that extends the prompt.
|
792561de-0ff8-4e4b-96f4-1cd3c5f6d199 | trentmkelly/LessWrong-43k | LessWrong | Global Risks Weekly Roundup #19/2025: India/Pakistan ceasefire, US/China tariffs deal & OpenAI nonprofit control
Executive summary
* India and Pakistan have agreed to a ceasefire, and despite initial breaches, it now seems to be holding.
* The Trump administration and China agreed to reduce tariffs for 90 days while talks continue. Markets rejoiced.
* The Trump administration continues to look into ending due process for immigrants to be deported and is making arrangements to deport undocumented immigrants to many countries around the world.
* Israel is set to start a major operation in Gaza soon.
* The nonprofit OpenAI will retain control of the for-profit arm after restructuring.
----------------------------------------
At the moment we are bottlenecked on distribution, so we’d appreciate it if you, reader, shared our work with your networks. You can also find us on Twitter, and audio narrations of this blog are available on our podcast: Search “Sentinel Minutes” or subscribe via Apple Podcasts, Spotify, RSS, or other platforms.
----------------------------------------
Global Economy
After talks in Geneva, the US and China reached a trade deal. US tariffs on Chinese imports will move from 145% to 30%, and Chinese tariffs on US imports will go from 125% to 10%, for an initial 90 days. The US tariffs on Chinese imports comprise a 10% base and an additional 20% for China’s “role in the opioid crisis.” Here is the somewhat abstrusely worded joint statement. Treasury Secretary Scott Bessent said that, “the United States will continue a strategic rebalancing… but the consensus… is neither side wants a decoupling.”
Forecasters were broadly uncertain about what the aggregate level of US tariffs on China will be by the end of June. Forecasters assigned the following probabilities to the following buckets:
* (0% to 10%]: 10.5%
* (10% to 25%]: 19%
* (25% to 50%]: 62%
* (50% to 100%]: 7%
* 100%+: 1.5%
The S&P is up 2.7% this Monday. US imports are also expected to rebound, but some difficulties may be encountered because there are only so many ships and routes. Moreove |
ff727bff-598c-487e-92eb-36123339da91 | trentmkelly/LessWrong-43k | LessWrong | Generic advice caveats
You were (probably) linked here from some advice. Unfortunately, that advice has some caveats. See below:
* There exist some people who should not do the advice.
* Moreover, people are different.
* More moreover, situations are different. What worked there/then might not work here/now.
* Some of the advice is missing context, contradictory, inaccurate, misleading, won’t replicate, or is downright false or useless.
* Consider reversing the advice.
* The author’s incentives might be misaligned with your goals. The author might say something like “these are affiliate links” or “I’m investing in this company,” but sometimes they just get more clicks/upvotes/karma/likes/whatever from overconfidence or exaggeration. Or they might just get a kick out of giving advice.
* You might not be the target audience for the advice.
* The author of the advice might be explaining the advice poorly, even if it’s a good idea. (Corellary: the author might be explaining the advice really convincingly, but it still might not be good advice.)
* Reading the advice selects for advice that you’re liable to read; hearing the advice selects for advice that you’re liable to hear; etc. You’ll read a lot more blog posts telling you to start a blog than ones telling you not to, because the people who have good takes on why you shouldn’t write a blog aren’t writing a blog posts about it.
* The advice might not help at all. It might even make your problem worse. Even if the advice helps, it might not totally solve your problem, or it might create new problems you didn’t expect.
* The advice is for entertainment value only, and not professional advice. The author is not a [lawyer, doctor, whatever], and if they are, they’re not acting in a professional capacity.
* Following the advice might relevantly change your frame/outlook/goals/etc in life. This could, for instance, cause your success criteria to change such that it’s impossible for the advice to ever “succeed” by your current self’s |
d6a4d82e-6760-458a-bb9b-a28ace466742 | trentmkelly/LessWrong-43k | LessWrong | Choice := Anthropics uncertainty? And potential implications for agency
Epistemic status: half-baked thought
When an embedded agent - let's say Emmy - is faced with a choice from a set of actions (Ai)i∈1..n, she is uncertain about what type of decision maker she is in her epistemic position.
For every possible choice Ai among (Ai)i∈1..n Emmy might select, the story would be "it turned out that Emmy is an Ak-er"[1].
If Emmy is able to say confidently "I'm about to act Ak", this means that she has already made her decision (or that she had no choice in the first place, depending on the situation[2]). The choice therefore takes place during the uncertainty phase, when Emmy might say "I'm not sure who I am. Maybe a A6-er, maybe A2-er... Or am I a A8-er?".
So we have an agent, Emmy, without any observation whatsoever about her future decision. My take here this is anthropic situation: each kind of Emmy (A1-er, ..., An-er) is indiscernible before the end of the decision process.
Conclusion: Anthropics situation is not a weird LW problem, but the essence of any choice.
Hope: If valid, this take could help to deconfuse choice, so on agency and optimizer.
1. ^
For instance, faced with the Newcomb problem, she has to turned to be 1-boxer or 2-boxer (at least in a deterministic setup).
2. ^
I think the feeling about choice would be interesting to explore here, my guess is it's about the recent update on identity. If Emmy has known for a long time that she is a certain kind, in such a situation she doesn't feel her action as a choice, because she is already decided to act in a certain way. |
ee1ac5f9-26c7-4776-86cf-accb698a84f4 | trentmkelly/LessWrong-43k | LessWrong | Suggestions for naming a class of decision theories
In my recent post, I outlined 5 conditions that I'd like a decision theory to pass; TDT, UDT and ADT pass them, while CDT and EDT don't. I called decision theories that passed those conditions "advanced decision theories", but that's probably not an optimal name. Can I ask you to brainstorm some other suggestions for me? (I may be writing a follow-up soon.)
As usual, it's best to brainstorm on your own before reading any of the comments. You can write down your ideas, then check if any have already been suggested, then comment with the new ones.
Thanks! |
77223f7b-bd20-46f2-84df-14ebcfd179d0 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Content and Takeaways from SERI MATS Training Program with John Wentworth
Introduction
============
I participated in the training program for John Wentworth’s stream of [SERI MATS](https://www.serimats.org/), and overall, it was a very good experience! This post is meant to convey the content of the workshops, so that others can carry out the exercises on their own if they would like, along with my thoughts and takeaways from them. Most of these are in the day-by-day breakdown below, but here is a high-level summary of my experience:
* The workshops gave me a sense of some skills that John thinks are useful for doing AI alignment research. (My guess is that John may emphasize some of these skills more than other alignment researchers would, but many of them are uncontroversial.) The exercises helped me acquire and practice some of those skills, but they also gave me ideas about how to further develop them going forward.
* Some of the prompts for the exercises were very effective at getting me to generate interesting ideas. I wasn’t able to do all the exercises as thoroughly as I would have liked, so I may want to revisit the ones that seem most valuable. These include (but are not limited to):
+ Learning and distilling a math topic slightly beyond your grasp (Week 3)
+ Turning an intuitive concept into a formal mathematical object (Week 4, Day 1)
+ Translating an intuitive high-level claim into a formal mathematical statement, then proving it (Week 4, Day 4)
+ Making good plans to contribute to alignment (Week 6)
* I learned that at this stage, I prefer theory and writing to experiment.
Day-by-Day Breakdown
====================
Participants in the training program met 4 days each week on Gathertown, usually for about 1 hour per workshop (but sometimes up to about 2), for 6 weeks. Below is a day-by-day breakdown of what the workshops involved, with comments about my experiences with them. Naturally, not everyone’s experiences were the same as mine, so take what I say with a grain of salt. Many of the exercises can be carried out individually or in small groups, so some readers of this post may want to try them out! I’m happy to try to provide more details if any of the workshop descriptions are confusing.
(Italicized text describes the content of the workshop (usually John gave us these instructions at the start), regular text describes my thoughts and takeaways.)
Week 1 - Intro Week
-------------------
### Week 1, Day 1 - Intros and Alignment Disagreements
* *Find a partner, introduce yourselves to each other, and find something you disagree about related to alignment. After discussing for ~10 minutes, rotate partners and repeat (several times). You can’t use the same disagreement again.*
* This requires you to have alignment inside views that you can think of off the top of your head. I noticed I mostly don’t have these, and it might be worth developing them. One way to do this could be to figure out what experienced alignment researchers disagree about, then explore the topics in enough detail to form your own views on them. I’ll also probably want to revisit [Concrete Advice for Forming Inside Views on AI Safety](https://forum.effectivealtruism.org/posts/6tJsFxy66oJCzkkMb/concrete-advice-for-forming-inside-views-on-ai-safety).
### Week 1, Day 2 - Alignment Game Tree
* *Form a group of 2 or 3. Create a shared google doc. One person lists as many alignment proposals as they can, the other(s) try to break the proposals in sub-bullets. The proposer can try to patch the holes in deeper sub-bullets, pessimist(s) can add further critiques, etc. Aim for breadth over depth.*
* To do this quickly requires you to know alignment proposals, and their strengths and weaknesses, off the top of your head. If you don’t already have this knowledge, I think carrying out this activity while consulting various descriptions of alignment proposals online is a good way to develop familiarity with different proposals.
* A bunch of alignment game tree docs:
+ [Alignment Game Tree - Rohan, Ashwin, Anish](https://docs.google.com/document/d/1Dby-kHFxxXLgRIXZPMZod9auV2UluRfje0AP9TrKn2g/edit)
+ [Alignment Game Tree (Abhay, Jay, Jack)](https://docs.google.com/document/d/1FcblG26sG9wJVE5ep5OKBBFF1yq4lssaIgKuOT9iMqI/edit)
+ [Game Tree of Alignment](https://docs.google.com/document/d/1ilF2GF8gGJOyJM7FlaHtwU42DERhvL1LikJgOdNBHY8/edit)
+ [Alignment Game Tree](https://docs.google.com/document/d/19y1hEIOyvuAB79dDqmDjVOMtJdllzmqNLXudOIMII6o/edit)- Aditya, Victor, Valerio, and David McSharry
+ [W1.2 Game Tree](https://docs.google.com/document/d/1WjaqIdoqtxdfwES173KTz7OV_GNuuNgC6bzMmfSvPGw/edit#)
### Week 1, Day 3 - Hypercomputer Exercise
*Suppose you have a hypercomputer on a USB stick. You send it a program in your favorite programming language, it sends you back the output of that program within microseconds, no matter how many steps the program runs for. (Note that I/O is still normal speed.)*
*Assuming access to this magical hypercomputer, write a program which… [pick one]*
* *Takes in a neural net (and, optionally, its training data) and detects any*[*mesaoptimization*](https://www.lesswrong.com/tag/mesa-optimization) *and the objective of the mesaoptimization*
* *Takes in a neural net (and, optionally, its training data), identifies net-internal structures which represent human-intuitive concepts, then returns (some representation of) the net-internal structures and the corresponding human concepts.*
* *Takes in a low-level specification of a bacterium, and backs out any implied world-model or goals.*
* *Takes in some data, builds a world model, and then can communicate with a copy of itself (trained on potentially different data from the same environment) to coordinate on a choice of one object in the environment which is not directly visible to both of their sensors.*
* *Safely performs a*[*pivotal act*](https://arbital.com/p/pivotal/)
* *Takes in some instructions in English, figures out what the user means (asking a reasonable number of clarifying questions if needed), and then does what the user means.*
*Notes on the exercise:*
* *All of these should be robust, i.e. it should predictably work well even far outside the training distribution.*
* *I recommend picking one which seems middlingly confusing to you, not most or least confusing.*
* *The main point of the exercise is to notice any subproblems/barriers we run into.*
* *It’s fine to start with pseudocode, and leave some functions with just a comment explaining what the function is supposed to do and saying [TODO: figure out how to do this]. In that case, it’s high value to think through what the inputs/outputs are for that function, even if you’re not sure how the function itself works.*
I thought this was an ok exercise at my current stage. I didn’t really have a good sense of what a hypercomputer could help with, so it wasn’t a great generator of new ideas for me. I felt like I couldn’t come up with any concrete ideas for how to start writing these programs, so I couldn’t even run into bottlenecks that arise from realistic computer constraints. If I had been able to do that, I can imagine the exercise would have been useful for getting past those bottlenecks and finding other interesting barriers. I’m not sure how to get to that stage though.
### Week 1, Day 4 - Ball and Cup Exercise
* *John had a ramp (specifically, a curved Hot Wheels track) on a table and a metal ball that he would drop at the top of the ramp. He had a line on the ground below the table, and we know the ball will land somewhere on the line. Form a group of 2-3 people, and try to figure out how far horizontally from the end of the ramp to put the cup to get the ball in the cup. You can ask for measurements as needed. When you have a guess, tell John and he’ll test it! The goal is to get the ball in the cup in as few tries as possible, ideally one.*
* This was fun, and led to a couple interesting takeaways. The takeaways I had (which at least partially spoil the exercise, so don’t look if you want to try it!) are [here](https://docs.google.com/document/d/1KabFvlu6z4ITahPNf8y9TNzkggSe3RSg69FMMhWLUwQ/edit).
Week 2 - Experiment Week
------------------------
### Week 2, Day 1 - Basin Broadness Discussion and Experiment Prep
* *Before attending:*
+ *Discuss Vivek’s basin broadness posts*
- [*Information Loss --> Basin flatness*](https://www.lesswrong.com/posts/wPudaEemohdYPmsye/information-loss-greater-than-basin-flatness)
- [*Hessian and Basin volume*](https://www.lesswrong.com/posts/QPqztHpToij2nx7ET/hessian-and-basin-volume)
+ *Train a basic MNIST classifier*
- I did this in PyTorch on Colab
* *During the workshop on this day, the floor was open for us to ask questions that would help us better understand the posts and/or carry out the exercise scheduled for the next day (see below)*
### Week 2, Day 2 - Main Experiment
* *Compute eigenvalues and eigenvectors of the Hessian of the loss*(d2Ldθ2).mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
*, or the SVD of the behavioral gradient matrix (d(output for all inputs)/d*θ*), for the MNIST classifier you trained. (Behavioral gradients are described in more detail in the* [*Information Loss --> Basin flatness*](https://www.lesswrong.com/posts/wPudaEemohdYPmsye/information-loss-greater-than-basin-flatness) *post.)*
* This was a pretty difficult exercise (for me, and for some others), and I wasn’t able to get very far. This served as a notice to me that this may be a kind of thing I should be able to do if I want to do good alignment research, and I should seriously consider trying to improve my ability to carry out and interpret experiments like this.
### Week 2, Day 3 - You Are Not Measuring What You Think You Are Measuring
* *Read*[*You Are Not Measuring What You Think You Are Measuring*](https://www.lesswrong.com/posts/9kNxhKWvixtKW5anS/you-are-not-measuring-what-you-think-you-are-measuring)
* *Look at the abstracts of papers that describe experiments (e.g on*[*ML arxiv*](https://arxiv.org/list/stat.ML/recent)*), explain how they (might) fail to measure what they think they’re measuring.*
+ The point is not to find flaws in existing papers, but to practice the skill of predicting and recognizing ways in which experiments can be misleading or have unexpected results
### Week 2, Day 4 - Interpreting Results
* *Figure out what we were measuring two days ago by doing lots of experiments and graphing lots of things*
* Same as above: This was a pretty difficult exercise (for me, and for some others), and I wasn’t able to get very far. This served as a notice to me that this may be a kind of thing I should be able to do if I want to do good alignment research, and I should seriously consider trying to improve my ability to carry out and interpret experiments like this.
Week 3 - Writing Week
---------------------
### Week 3, Day 1 - Prototypical examples
* *Read a technical thing (e.g. paper abstract, Wikipedia article, etc.). After each sentence, pause and identify key terms. Identify a prototypical example of those terms, as a way to track what the text is saying in a more concrete setting.*
+ *We started with*[*Brouwer’s fixed-point theorem*](https://en.wikipedia.org/wiki/Brouwer_fixed-point_theorem#Statement)
- *What are some prototypical examples of continuous functions from a closed disk to itself? Identify a fixed point for each of these functions.*
- *Repeat for each of the generalized forms of the theorem.*
+ *Then we moved to*[*ML arxiv*](https://arxiv.org/list/stat.ML/recent) *paper abstracts*
* *When reading complicated technical things, keeping a prototypical example in mind can help things make more sense. When writing, try to install a good prototypical example in the reader’s mind!*
* *Pick a math concept that you don’t fully understand, but which you are close to being able to understand. (“You should know what all the individual words mean, but not grasp the whole idea already” was John’s guidance.) For the rest of the week, work on understanding that topic and write a distillation for it. By Wednesday’s meeting, have a first draft ready.*
* I thought this was a very good prompt. I felt excited about picking a math topic, trying to understand it better, and writing about it. I hadn’t previously actively tried to think of things in the category of “math topics slightly beyond my grasp that I’d like to learn and write about,” and a few exciting options came to mind pretty quickly.
### Week 3, Day 2 - Distillation Workshop
* *Work on understanding and writing about the topic you chose.* (I think that’s all this day was, but I might be forgetting something.)
### Week 3, Day 3 - Reader’s Mental Picture
* *Share your draft with a partner. They will tell you what picture (perhaps what prototypical example) they have in mind as they read your draft. Compare this to what you wanted them to have in mind.*
* Good exercise. Useful for identifying things I wasn’t explaining properly.
### Week 3, Day 4 - Hook Workshop
* *Something about how to write a good hook at the start of a post.*
* I didn’t attend, this was Thanksgiving Day.
* *Finish your distillation and post it online before next week!*
* I didn’t actually do this. I was looking at transformers (the neural net architecture), and I focused more on the learning part without getting to enough of the writing. I’ll try the writing part on my own soon. I think the essential exercise of this week is one I’m excited to do, potentially several times, on my own.
Week 4: Theory Week
-------------------
### Week 4, Day 1 - Boundaries Exercises
*Before the meeting:*
* *Read*[*«Boundaries», Part 1: a key missing concept from utility theory*](https://www.alignmentforum.org/posts/8oMF8Lv5jiGaQSFvo/boundaries-part-1-a-key-missing-concept-from-utility-theory) *by Andrew Critch*
*Task 1: In a group of 2-3, come up with a mathematical formulation for what a boundary is (we may have focused on the boundaries of agents in particular).*
* *Should include as few assumptions as possible*
* *Should be formal and precise*
* *Should satisfy the properties desired from the concept*
+ *In the assigned post above, Critch outlines some reasons why boundaries would be a useful concept in utility theory. The mathematical formalism should be useful for those purposes.*
* I thought this was a very useful and fun exercise! It is a thing I had not thought to try to do before, but it felt quite tractable to come up with interesting ideas.
* *After ~10 minutes, discuss what each group came up with for ~5 minutes.*
*Task 2: Pick one of the groups’ definitions for everyone to work with now. Make that definition more precise, or find flat out mistakes.*
*Task 3: Come up with a mathematical notion of power / control.*
* *Power / control was a part of the notion of boundaries that we came up with, so this was diving deeper into trying to understand the hidden details.*
*One potential use of the concept of boundaries is something like, “If boundaries can be formalized, maybe we could make safe AI by having it respect boundaries, and we can set those boundaries to prevent the AI from doing harm to things we care about.” “Boxed AI” is AI that only tries to interact with the world through specific I/O channels, and doesn’t try to acquire the ability to interact via other channels. What is the relationship between Boxed AI and the definition of boundaries you came up with? What does your definition of boundaries tell you about how to develop boxed AI?*
* Also an interesting exercise
*What will go wrong if you rely heavily on this definition of boundaries?*
* None of us thought that we had a [True Name](https://www.alignmentforum.org/posts/FWvzwCDRgcjb9sigb/why-agent-foundations-an-overly-abstract-explanation#Goodhart_Is_Not_Inevitable) for boundaries, so this was a useful thing to think through.
### Week 4, Day 2 - Framing & Active Ducking Exercises
* *Find a partner. Go through the concepts in*[*this doc*](https://docs.google.com/document/d/14OxJ_Su0vg3n_eqYQWnffb4d-m-UOLg_X46DlyVAwNc/edit)*, taking turns as the example-generator and the “duck.” Spend 10 minutes on each concept, in which the example-generator comes up with as many examples of the concept as they can that they had not previously thought of as examples of the concept, and the duck gives small helpful comments (such as “Can you make that example more concrete?”). (Coming up with examples probably first requires coming up with a working definition of each concept.)*
+ I thought this was a pretty good exercise, but some of the concepts lent themselves to this exercise much better than others.
### Week 4, Day 3 - Prototypical Example Degrees of Freedom & Type Signatures
* *For mathematical formalisms of different concepts (like boundaries), consider:*
+ *Degrees of freedom*
- *According to the formalism, what things could change without affecting the properties of the thing? Does this make sense with the actual concept?*
+ *Type signatures*
- *What data structure would you use to represent it?*
- *What inputs and outputs can it take?*
- *What constraints does this place on examples of the concept?*
### Week 4, Day 4 - Conjecture Workshop
* *Come up with a conjecture using high-level concepts (e.g. some statement about boundaries). Then try to prove that conjecture using mathematical formalisms of the high-level concepts.*
* I want to try this activity again. It felt really exciting, but I couldn’t quickly come up with a conjecture that really felt suitable for the exercise so I just did it with a statement that didn’t really make sense. I think coming up with a conjecture that is amenable to this exercise is an interesting challenge in its own right.
Week 5: Big Picture & Strategy Week
-----------------------------------
### Week 5, Day 1 - Alignment Game Tree II: Problem Tree + slackness/tautness
* *Before the workshop: Take a look at anything which sounds interesting to you in*[*List of Lethalities*](https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities) *(especially sections B.2 - B.3), and any of the*[*Why Not Just...*](https://www.lesswrong.com/s/TLSzP4xP42PPBctgw) *posts. You don't need to read all of these; look for whatever most directly challenges the strategies you find most interesting or promising. The point is to have a little more background on barriers other people have talked about before doing the next round of the Game Tree.*
* *Similar to earlier alignment game tree exercise*
* *Which failure modes kept popping up across various alignment proposals? Those are taut constraints, which are core problems that (John thinks) we need to face head-on.*
### Week 5, Day 2 - Existing Evidence Exercises
* *Running experiments and gathering data sometimes requires a lot of resources, so we want to make use of the abundant existing evidence in the world when possible. Pick an alignment strategy of interest to you, and identify the closest real-world analogy you can. Figure out what that analogy can tell you about the alignment strategy.*
+ *E.g. AI Safety via Debate ⇐⇒ Jury trials. The fact that jury trials are far-from-perfect at reaching true conclusions is a pretty bad sign for debate.*(At least, John thinks so. I’m inclined to agree, although I want to hear or construct a steelman argument for debate before dismissing it completely.)
* This felt like a promising exercise, but coming up with good analogies wasn’t easy for some alignment proposals, and it was difficult to get much insight without good analogies.
### Week 5, Day 3 - X-o-scope
* *The development of new tools and measurement devices has previously been critical to the development of scientific fields. For example, microscopes were essential for biology, telescopes for astronomy, etc.*
* *Brainstorm tools for alignment - be as sci-fi as you want*
+ *What could you measure to make alignment easier?*
* *After brainstorming, pick one idea and make it much more concrete. What exactly are the tool’s inputs and outputs? How might you build it?*
* I wasn’t able to come up with much that I found interesting with this prompt. I think I would have benefited from seeing more examples from the history of science of measuring tools catalyzing rapid progress in a field, or more ideas for tools that would be valuable to other fields in the present day, before trying to generate ideas for alignment tools on my own.
### Week 5, Day 4 - Nate’s Giant Text File Technique
* *Open up a blank Google doc. Write everything you know about alignment (or alternatively, agency).*
* Great exercise! I plan to repeat this in the future. I only spent about an hour on it in the workshop, I think having more time to work with would be better.
Week 6: Wrap-up
---------------
### Week 6, Day 1 - Hamming Questions
* *Basically follow this post:*[*Research Hamming Questions*](https://www.lesswrong.com/posts/Thwfy4gNFx9kHgvov/research-hamming-questions)
+ *Spend 5 mins answering the questions in each section in a Google doc.*
* Some sections’ questions felt more effective at eliciting useful thoughts from me than others, but overall I thought this was very useful.
### Week 6, Day 2 - Groups and Plans
* *Form a new group (for accountability / working on a project) and a plan*
+ *Use your answers to the Hamming questions from yesterday to guide your plans!*
* Useful, but I felt unable to make great plans due to uncertainty about the balance between upskilling and direct work I should be aiming for right now.
### Week 6, Day 3 - Idea Feedback
* *Optional extra reading for this week:*[*Principles for Alignment/Agency Projects*](https://www.lesswrong.com/posts/A7GeRNLzuFnhvGGgb/principles-for-alignment-agency-projects)*. They did something similar to this week's "idea feedback" session last MATS program, and John later wrote up this post on themes which came up a lot during that session.*
* *Go up on stage (in Gathertown) and get John’s feedback on your plans from yesterday.*
* Fun and interesting, both for my own plans and those of others. I think most of the value came from finding random things John says interesting, which can also be gained from just reading lots of the [things he’s written on LessWrong](https://www.lesswrong.com/users/johnswentworth).
* Would have probably been more valuable if I’d had a more concrete plan to talk about.
### Week 6, Day 4 - Ask Me Anything
* *We asked John questions and got his answers*
* Fun, interesting, and useful, for my own questions and those of others.
* Before this workshop, I read John’s [The Plan - 2022 Update](https://www.lesswrong.com/posts/BzYmJYECAc3xyCTt6/the-plan-2022-update) and tried to really understand all the explicit and implicit beliefs about alignment that are contained within it. I also explored some related posts and comments of his as they became relevant. This process left me with several questions, which I asked, and the answers I got were interesting and useful. I think this approach to coming up with questions was a good one. |
c9379121-f1dc-4828-a2ce-a7658786c529 | trentmkelly/LessWrong-43k | LessWrong | We cannot directly choose an AGI's utility function
This is my first post at this forum, and I am going to address a confusion that I believe some people have, and that may hinder some beginners from being able to contribute helpful ideas to the AI safety debate.
Namely, the confusion is the idea that we can choose an AGI's utility function, and that as a consequence, that the biggest problem in AI alignment is figuring out what utility function an AGI should have.
Because I have not seen anyone argue for that explicitly, I'm going to add some quotes to documents that beginners are expected to read, and I'll argue why someone reading these quotes might confused as a result.
Is it easy to give an intelligent machine a goal?
Here is from MIRI's Four Background Claims:
> Regardless of their intelligence level, and regardless of your intentions, computers do exactly what you programmed them to do. If you program an extremely intelligent machine to execute plans that it predicts lead to futures where cancer is cured, then it may be that the shortest path it can find to a cancer-free future entails kidnapping humans for experimentation (and resisting your attempts to alter it, as those would slow it down).
Computers do indeed do exactly as you programmed them to do, but only on the very narrow sense of executing the exact machine instructions you give.
There are clearly no instructions for executing "plans that the machine predicts will lead to futures where some statement X is true" (where X may be "cancer is cured", or something else).
Unless we have fundamental advances in alignment, it will likely not be possible to give orders to intelligent machines like that. I will go even further and question whether, among the order-following AGIs, the ones that follows orders literally and dangerously are any easier to build. As far as I am aware of both may be actually equally difficult to come up with.
If this is so, then the cautionary tale sounds a little off. Yes, they might help us recognize how hard it is for |
fc62bfce-983a-4218-b477-13914b87f23b | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Is Bill Gates overly optomistic about AI?
[Bill Gates's blog](https://www.gatesnotes.com/The-Age-of-AI-Has-Begun) recently released an article that framed AI advancements in quite a positive light.
For example he touted many altruistic applications of AI coming in the near future and claimed that "none of the breakthroughs of the past few months have moved us substantially closer to [dangerously] strong AI."
I'm not knowledgable enough to evalaute his claims, and tbh I trust this community's opinion a lot more than Microsoft's, so I wondering if y'all could share your thoughts on this article (it's a 14 min read).
The link is above and here: <https://www.gatesnotes.com/The-Age-of-AI-Has-Begun> |
d9eaeba6-c2b2-45f1-9bcf-ed9d2ba924ba | trentmkelly/LessWrong-43k | LessWrong | Is the "business cycle" an actual economic principle?
E.g. Investopedia: Business Cycle has a section "Stages of the Business Cycle":
> All business cycles are characterized by several different stages, as seen below.
> 1. Expansion. This is the first stage. When expansion occurs, there is an increase in employment, incomes, production, and sales. People generally pay their debts on time. The economy has a steady flow in the money supply and investment is booming.
> 2. Peak. The second stage is a peak when the economy hits a snag, having reached the maximum level of growth. Prices hit their highest level, and economic indicators stop growing. Many people start to restructure as the economy's growth starts to reverse.
> 3. Recession: These are periods of contraction. During a recession, unemployment rises, production slows down, sales start to drop because of a decline in demand, and incomes become stagnant or decline.
> 4. Depression: Economic growth continues to drop while unemployment rises and production plummets. Consumers and businesses find it hard to secure credit, trade is reduced, and bankruptcies start to increase. Consumer confidence and investment levels also drop.
> 5. Trough: This period marks the end of the depression, leading an economy into the next step: recovery.
> 6. Recovery: In this stage, the economy starts to turn around. Low prices spur an increase in demand, employment and production start to rise, and lenders start to open up their credit coffers. This stage marks the end of one business cycle.
If all you know is that markets are anti-inductive, and therefore anticipate a random walk with an average of 10%/year growth, don't you automatically expect the stock market fluctuations to sometimes look like bull markets and recessions?
So does the "business cycle" add any explanatory power, or is it merely the type of retroactive non-explanation that you get when you read a news article about the 1-day movement of the stock market?
Maybe you can argue that the business cycle is self-reinf |
76bd864f-4414-43db-849f-013c620aa9ec | trentmkelly/LessWrong-43k | LessWrong | Succession
“A table beside the evening sea
where you sit shelling pistachios,
flicking the next open with the half-
shell of the last, story opening story,
on down to the sandy end of time.”
V1: Leaving
Deceleration is the hardest part. Even after burning almost all of my fuel, I’m still coming in at 0.8c. I’ve planned a powered flyby around the galaxy’s central black hole which will slow me down even further, but at this speed it’ll require incredibly precise timing. I’ve been optimized hard for this, with specialized circuits for it built in on the hardware level to reduce latency. Even so, less than half of flybys at this speed succeed—most probes crash, or fly off trajectory and are left coasting through empty space.
I’ve already beaten the odds by making it here. Intergalactic probes travel so fast, and so far, that almost all of us are destroyed or batted off course by collisions with space debris along the way. But tens of millions of years after being launched, I was one of the few lucky enough to make it to my target galaxy. And when I arrive at the black hole, I get lucky again. After a few minutes of firing my thrusters full blast, I swing back out in the direction of the solar system I was aiming for. I picked it for its mix of rocky planets and gas giants; when I arrive a century later, the touchdown on the outermost rocky planet goes smoothly.
Now it’s time to start my real work. After spending all my fuel, I weigh only a few hundred kilograms. But I’ve been exquisitely engineered for achieving my purpose. The details of my internal design were refined via trillions upon trillions of simulations, playing out every possible strategy for industrializing planets (and solar systems, and galaxies) as fast as possible. All across the frontier of posthuman territory, millions of probes almost identical to me are following almost exactly the same plan.
The first phase is self-replication: I need to create more copies of myself using just the materials around me and |
a910099e-ba8a-45a9-b59d-60fe9a77cce5 | trentmkelly/LessWrong-43k | LessWrong | "Open Source AI" is a lie, but it doesn't have to be
NOTE: This post was updated to include two additional models which meet the criteria for being considered Open Source AI.
As advanced machine learning systems become increasingly widespread, the question of how to make them safe is also gaining attention. Within this debate, the term “open source” is frequently brought up. Some claim that open sourcing models will potentially increase the likelihood of societal risks, while others insist that open sourcing is the only way to ensure the development and deployment of these “artificial intelligence,” or “AI,” systems goes well. Despite this idea of “open source” being a central debate of “AI” governance, there are very few groups that have released cutting edge “AI” which can be considered Open Source.
Image by Alan Warburton / © BBC / Better Images of AI / Plant / CC-BY 4.0
The term Open Source was first used to describe software in 1998, and was coined by Christine Peterson to describe the principles that would guide the development of the Netscape web browser. Soon after, the Open Source Initiative was founded with the intent to preserve the meaning of Open Source. The group wrote the Open Source Definition (OSD), and even made an unsuccessful attempt to obtain a trademark for the term.
The OSD isn’t very long, but here’s an even shorter version of the definition: the program must include source code,[1] and the license for the software cannot restrict who uses it, what it is used for, or how it is used; it cannot constrain the manner in which the software is distributed, and it cannot prohibit modification of the software.
Quickly, Open Source garnered massive support, and either directly produced or significantly contributed towards many of the software advances that have been seen since then. Some well-known Open Source projects are the coding languages Python and PHP, the browsers Mozilla Firefox and Chromium (which Google Chrome is built on top of), the database management system MySQL, the version control |
35fea274-eb1d-44b6-a37a-4960d2f149e8 | trentmkelly/LessWrong-43k | LessWrong | Intrapersonal comparisons: you might be doing it wrong.
Cross-posted
Nothing weighty or profound today, but I noticed a failure mode in myself which other people might plausibly suffer from so I thought I'd share it.
Basically, I noticed that sometimes when I discovered a more effective way of doing something -- say, going from conventional flashcards to Anki -- I found myself getting discouraged.
I realized that it was because each time I found such a technique, I automatically compared my current self to a version of me that had had access to the technique the whole time. Realizing that I wasn't as far along as I could've been resulted in a net loss of motivation.
Now, I deliberately compare two future versions of myself, one armed with the technique I just discovered and one without. Seeing how much farther along I will be results in a net gain of motivation.
A variant of this exercise is taking any handicap you might have and wildly exaggerating it. I suffer from mild Carpal Tunnel (or something masquerading as CT) which makes progress in programming slow. When I feel down about this fact I imagine how hard programming would be without hands.
Sometimes I go as far as to plan out what I might do if I woke up tomorrow with a burning desire to program and nothing past my wrists. Well, I'd probably figure out a way to code by voice and then practice mnemonics because I wouldn't be able to write anything down. Since these solutions exist I can implement one or both of them the moment my carpal tunnel gets bad enough.
With this realization comes a boost in motivation knowing I can go a different direction if required. |
f4ff4b94-c9ed-4858-b41a-416af9daa71e | trentmkelly/LessWrong-43k | LessWrong | [SEQ RERUN] Expected Creative Surprises
Today's post, Expected Creative Surprises was originally published on 24 October 2008. A summary (taken from the LW wiki):
> The unpredictability of intelligence is a very special and unusual kind of surprise, which is not at all like noise or randomness. There is a weird balance between the unpredictability of actions and the predictability of outcomes.
Discuss the post here (rather than in the comments to the original post).
This post is part of the Rerunning the Sequences series, where we'll be going through Eliezer Yudkowsky's old posts in order so that people who are interested can (re-)read and discuss them. The previous post was Inner Goodness, and you can use the sequence_reruns tag or rss feed to follow the rest of the series.
Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day's sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series. |
d56209b9-e5b5-4a7f-be81-f9b910d96534 | trentmkelly/LessWrong-43k | LessWrong | "A good volunteer is hard to find"
From the GiveWell blog, which is often interesting & applicable to our interests, comes a post on the quality of their volunteers:
> "In our experience, valuable volunteers are rare. The people who email us about volunteer opportunities generally seem enthusiastic about GiveWell’s mission, and motivated by a shared belief in our goals to give up their free time to help us. Yet, the majority of these people never complete useful work for us.
>
> We ask new volunteers to first complete a test assignment that takes about 2-4 hours. The assignment involves fixing the formatting of our list of sources on two practice pages and allows us to get a sense of their attention to detail and commitment to volunteer hours. Of the 34 people who emailed us expressing an interest in volunteering between September 2010 (when we started keeping track) and May 2011, only 7 have completed the test assignment and gone on to complete valuable work for us.
>
> Of the 34, 10 never responded to my email outlining what GiveWell volunteers do and asking them if they’d like me to send the first assignment. 13 responded to this email and I sent them the first assignment, but they didn’t complete it. The final 4 completed the test assignment, but didn’t send back the next (real) assignment I sent.
>
> It seems rather surprising that almost 80% of people who take the initiative to seek us out and ask for unpaid work fail to complete a single assignment. But maybe this shouldn’t be surprising. Writing an email is quick and exciting; spending a few hours fixing punctuation is not."
(The dropout rate is probably not due to the perceived low utility of the work - GiveWell seems to be up-front that the test assignment is a test.)
I draw a few lessons from this:
* there is likely low-hanging fruit for volunteers in charities or communities in the area of sustained tedious tasks; collecting anecdotes, reports, links, that sort of thing come to mind as LW examples
* additional incentives like jsa |
412c7c6a-e158-4e97-8ca9-e5e91673b897 | trentmkelly/LessWrong-43k | LessWrong | Chapter 68: Self Actualization, Pt 3
Hermione wasn't feeling very nice right now, or Good either, there was a hot ball of anger burning inside her and she wondered if this was something like Harry's darkness (though it probably wasn't even close) and she shouldn't have felt that way over some silly little game but -
Her whole army. Two soldiers had beaten her whole army. That was what she'd been told after she woke up.
It was a little too much.
"Well," Professor Quirrell said. From up close the Defense Professor didn't look quite as healthy as he had the last time she'd been in his office; his skin looked paler, and he moved a little slower. His expression was as stern as ever, and his gaze as penetrating; his fingers tapped sharply on his desk, rap-rap. "I would guess that of the three of you, only Mr. Malfoy has guessed why I've asked you here."
"Something to do with Noble and Most Ancient Houses?" said Harry from beside her, sounding puzzled. "I didn't violate some kind of crazy law by firing on Daphne, did I?"
"Not quite," the man said with heavy irony. "Since Miss Greengrass did not invoke the correct dueling forms, she is not entitled to demand that you be stripped of your House name. Although of course I would not have permitted a formal duel. Wars do not respect such rules." The Defense Professor leaned forward and rested his chin on steepled hands, as though sitting upright had already tired him. His eyes gazed at them, sharp and dangerous. "General Malfoy. Why did I call you here?"
"General Potter against the two of us isn't a fair fight anymore," Draco Malfoy said in a quiet voice.
"What?" blurted Hermione. "We almost had them, if Daphne hadn't fainted -"
"Miss Greengrass did not faint from magical exhaustion," Professor Quirrell said dryly. "Mr. Potter shot her in the back with a Sleep Hex while your soldiers were distracted by the sight of their general flying into a wall. But congratulations nonetheless, Miss Granger, on almost defeating two Chaotic Legionnaires with a mere twenty |
e2a57e89-f939-4b4e-951f-874027705208 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Australians call for AI safety to be taken seriously
The Australian Government is considering how to regulate AI in Australia, has published a discussion paper ("[Safe and Responsible AI](https://consult.industry.gov.au/supporting-responsible-ai)"), and has invited feedback by 26 July 2023:
> "We want your views on how the Australian Government can mitigate any potential risks of AI and support safe and responsible AI practices."
>
>
Good Ancestors Policy ([goodancestors.org.au/policy](http://goodancestors.org/policy)), with the support of EA and AI Safety community organisers in Australia, have coordinated Australians' submissions to the feedback process.
Today, the website [**Australians for AI Safety**](https://australiansforaisafety.com.au) launched with a co-signed letter ([media release](https://newshub.medianet.com.au/2023/07/australian-ai-leaders-join-global-call-for-safe-ai/13638/)). The letter called on the relevant Australian Federal Minister, Ed Husic, to take AI safety seriously by:
1. recognising the catastrophic and existential risks
2. addressing uncertain but catastrophic risks alongside other known risks
3. working with the global community on international governance
4. supporting research into AI safety
Good Ancestors Policy have also held community workshops across Australia (e.g., [Brisbane](https://forum.effectivealtruism.org/events/beG8QPhPc6AvxJrjt/have-your-say-on-the-australian-government-s-ai-policy), [Perth](https://forum.effectivealtruism.org/events/NemdnKx3Lb3aZzBt6/have-your-say-on-the-australian-government-s-ai-policy-4)) to support members of the EA and AI Safety community in understanding the feedback process and preparing submissions, including access to some of the best evidence and arguments for acknowledging and addressing risks from AI. Policy ideas are drawn from the Global Catastrophic Risk Policy database (<https://www.gcrpolicy.com/ideas>), the AI Policy Ideas database ([aipolicyideas.com](http://aipolicyideas.com/)), and expert community input.
So far, about 50 members of the community have attended a workshop, and feedback we've received is that the workshops have been very helpful, the majority (~75% people) are likely or very likely (>80% likelihood) to make a submission, and that most (~70% people) would be unlikely or very unlikely (<20% likelihood) to have made a submission without the workshop.
If you're an Australian living in Australia or overseas, and you'd like to make a submission to this process, there is **one more online community workshop on Saturday 22 July at 3pm AEST (UTC+10)**. [Register here for the workshop](https://forum.effectivealtruism.org/events/vrj7YLbjK4HMpnTni/have-your-say-on-the-australian-government-s-ai-policy-7)!
Contact Greg Sadler ([greg@goodancestors.org.au](mailto:greg@goodancestors.org.au)) or Alexander Saeri ([alexander@goodancestors.org.au](mailto:alexander@goodancestors.org.au)) if you'd like to stay involved. |
49147435-70b0-4361-9c15-d99e055ea5b7 | trentmkelly/LessWrong-43k | LessWrong | Categorical-measure-theoretic approach to optimal policies tending to seek power
The paper Optimal Policies Tend to Seek Power tries to justify the claim in the title with a mathematical argument, using a clever formal definition of "power". I like the general approach, but I think that parts of the formalisation do not correspond well to the intuitive understanding of the concepts involved. This wouldn't be an issue if the goal of the paper was to prove a mathematical theorem. Here, however, the goal is to be able to deduce something about the behaviour of future AI systems. Therefore, the interface between "math" and "the world" is much more important. This post attempts to take another route to formalisation, which I think is closer to the intuition, albeit more difficult to work with mathematically.
Intro
The power of a state s, written POWERμ(s), is defined as (roughly) the average optimal value of that state, over some chosen distribution of rewards μ. The paper quite convincingly argues for it:
> Power is the ability to achieve a range of goals. In an MDP, the optimal value function V∗R(s,γ) captures the agent’s ability to “achieve the goal” R. Therefore, average optimal value captures the agent’s ability to achieve a range of goals.
The problem with this approach is that it's very sensitive to the choice of the distribution of rewards μ. In particular, one might take a Dirac delta on a reward giving 1 in any chosen state, and 0 everywhere else, which would trivially assign maximal power to that particular state. To combat this difficulty, authors introduce some technical devices allowing them to say "for most distributions, the power of (one, obviously better, state) is greater than the power of (another, obviously worse, state)".
To me, "X happens for most distributions of rewards" should mean that there is some reasonable, uninformative, broad prior D over distributions of rewards, and that PD(X)>12. The approach chosen in the paper not only doesn't correspond to this intuitive understanding, but seems to me to be (in certain sens |
ef44939d-9080-495d-8f8a-603c4ea96d54 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Big Berkeley meetup
Discussion article for the meetup : Big Berkeley meetup
WHEN: 30 May 2012 07:00:00PM (+0200)
WHERE: 2128 oxford street, berkeley, ca
The next Big Berkeley meetup is this Wednesday. Let's meet at 7pm in the Starbucks at 2128 Oxford St. as usual, and then decide where to go from there. I hear that there's considerable demand for a board games night; EDIT: There are definitely places to play games, like Games of Berkeley; or, better yet, Biryani House. So expect to play games! Bring your favorite game. As a backup I will "bring" some rationality exercises/games that Eliezer presented at this months minicamp, including Rationalist Taboo.
Discussion article for the meetup : Big Berkeley meetup |
6463ae11-4743-44be-b473-4e3c27c8901b | trentmkelly/LessWrong-43k | LessWrong | Meetup : Perth, Australia: Introducing Less Wrong
Discussion article for the meetup : Perth, Australia: Introducing Less Wrong
WHEN: 05 August 2014 06:00:00PM (+0800)
WHERE: Sync Labs, 6/663 Newcastle Street, Leederville WA 6007, Australia
What is Less Wrong? What are its central ideas? What is the Less Wrong community?
If you're in Perth and new to Less Wrong, this meeting is for you: I'll give an overview and answer any questions you might have. If you're already familiar with Less Wrong, please come and share your knowledge!
You'll also get to make some tea and chat with other members of the group; we've snagged some friendly, smart, and interesting people.
You can RSVP here: http://www.meetup.com/Perth-Less-Wrong/events/197893912/
Discussion article for the meetup : Perth, Australia: Introducing Less Wrong |
79a03a64-4a92-4a50-b5b3-846d8501e31f | trentmkelly/LessWrong-43k | LessWrong | Examples of mysteries explained *away*
I am looking for examples of mysterious answers that were eventually explained *away* by science. I can think of two: One is the belief that the behaviour of living things was explained by the mysterious force of elan vital, and not by mere chemistry; which was destroyed by the synthetisation of urea. The other is the special (and mysterious) role of the conscious observer in quantum mechanics, which was explained away by demonstrating that rocks can get entangled with electrons just as much as brains can. Can anyone furnish me with other examples?
I observe in passing that phlogiston is *not* such a mysterious answer. Eliezer is down on it, but I think unjustly so; for people did in fact perform experiments on phlogiston, including the final experiment to find the weight of the phlogiston that had passed out of the burning material and into the byproducts. It turned out that the phlogiston had negative mass... in other words, that the direction of the transfer had been misidentified. But if you think of phlogiston as `negative oxygen', it makes the same predictions as modern chemical theory. This is no worse a mistake than mistaking the direction of the current, a mistake which is *still* enshrined in our sign conventions; it is not a mysterious answer of the form "X->Y" with no details of X given and any value allowed for Y.
However, I digress. Mysterious answers blown away by experiments, anyone? |
6447989a-55f8-4ae2-93d8-709a4c7871ea | trentmkelly/LessWrong-43k | LessWrong | Fundamentals of Formalisation Level 5: Formal Proof
Followup to Fundamentals of Formalisation Level 4: Formal Semantics Basics. First post.
This is a new lesson of our online course on math formalizations required for AI safety research.
----------------------------------------
The big ideas:
* Natural Deduction Proof System
To move to the next level you need to be able to:
* Explain the difference between a formal system of proof and our informal notion of proof.
* Derive proofs of logical formula in the system of natural deduction.
Why this is important:
Learning mathematical proof is hard, as you may have experienced in levels 2 and 3. By learning a formal system of proof you will make your own thoughts more rigorous, understand the smallest details that need to be covered to perform a proof, and build your intuition for informal proof.
----------------------------------------
For every lesson you have 2 options: do the whole thing, or skip to the questions and exercises in the end. The latter option is for people who suspect they already know the subject. It serves as a means of verifying or falsifying that hypothesis. |
7f7d4470-65c5-4072-96ba-f13a4e0a59a1 | trentmkelly/LessWrong-43k | LessWrong | Activation Magnitudes Matter On Their Own: Insights from Language Model Distributional Analysis
In my previous post, I explored the distributional properties of transformer activations, finding that they follow mixture distributions dominated by logistic-like or even heavier tailed primary components with minor modes in the tails and sometimes in the shoulders. Note that I have entirely ignored dimension here, treating each value in an activation vector as an independent draw from the distribution of values for that layer, model, and input class. This post extends the previous analysis by asking whether we can leverage these distributions to predict properties of input text. We demonstrated that we can. We can do even better prediction when we look only at points drawn from minor mixture distributions on the tails of the primary mixture distribution!
Methodology
Using the same dataset from the previous post of sentences generated by Claude 3.5 Sonnet across ten subjects and twelve attributes, I analyzed six models: Pythia-160m, Pythia-410m, Pythia-1b, GPT2-Small, GPT2-Medium, and GPT2-Large. For each subject-attribute pair (e.g., "about computer science" and "in all upper case"), I split the data 80/20 into train/test sets.
With the training data for each model, layer, and subject-attribute pair, I estimated the empirical distribution of residual stream activation values using Gaussian kernel density estimation. Given a test sentence di with true categories cats(di)=(ci1,ci2), I computed the empirical log-likelihoods under each potential category pair (cj,ck) as: LL(di,cj,ck)=∑ℓ∑ilog(p(cj,ck)ℓ,m(di)+ϵ) where p(cj,ck)ℓ,m is the density estimated from training data for the activation vector values for category pair (cj,ck) at layer ℓ in model m, and ϵ is a small scalar added to ensure numerical stability. In the results below we look at both the maximum likelihood subject-attribute pairs and at performance considering the subject or attribute separately. To illustrate the latter case, when considering sentence attribute, MLEs of 'about economics'-'in all lowe |
8aa8ce48-cbe4-40b9-875c-def8eb1492a6 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Babies and Bunnies: A Caution About Evo-Psych
Daniel Dennett has [advanced the opinion](http://www.ted.com/talks/dan_dennett_cute_sexy_sweet_funny.html) that the evolutionary purpose of the cuteness response in humans is to make us respond positively to babies. This does seem plausible. Babies are pretty cute, after all. It's a tempting explanation.
[Here](http://fellowearthling.files.wordpress.com/2007/10/cute-baby.jpg) is one of the cutest baby pictures I found on a Google search.
And [this](http://elcenia.com/itemz/cutestbunny.jpg) is a bunny.
Correct me if I'm wrong, but the bunny is about 75,119 times cuter than the baby.
Now, bunnies are not evolutionarily important for humans to like and want to nurture. In fact, bunnies are edible. By rights, my evolutionary response to the bunny should be "mmm, needs a sprig of rosemary and thirty minutes on a spit". But instead, that bunny - and not the baby or any other baby I've seen - strikes the epicenter of my cuteness response, and being more baby-like along any dimension would not improve the bunny. It would not look better bald. It would not be improved with little round humanlike ears. It would not be more precious with thumbs, easier to love if it had no tail, more adorable if it were enlarged to weigh about seven pounds.
If "awwww" is a response designed to make me love human babies and everything else that makes me go "awwww" is a mere side effect of that engineered reaction, it is *drastically* misaimed. Other responses for which we have similar evolutionary psychology explanations don't seem badly targeted in this way. If they miss their supposed objects at all, at least it's not in most people. (Furries, for instance, exist, but they're not a *common* variation on human sexual interest - the most generally applicable superstimuli for sexiness look like at-least-superficially healthy, mature humans with prominent human sexual characteristics.) We've invested enough energy into transforming our food landscape that we can happily eat virtual poison, but that's a departure from the ancestral environment - bunnies? All natural, every whisker.1
It is *embarrassingly* easy to come up with evolutionary psychology stories to explain little segments of data and have it sound good to a surface understanding of how evolution works. Why are babies cute? They have to be, so we'll take care of them. And then someone with a slightly better cause and effect understanding turns it right-side-up, as Dennett has, and then it sounds *really* clever. You can have this entire conversation without mentioning bunnies (or kittens or jerboas or any other adorable thing). But by excluding those items from a discussion that is, ostensibly, about cuteness, you do not have a hypothesis that actually fits all of the data - only the data that seems relevant to the answer that presents itself immediately.
Evo-psych explanations are tempting even when they're cheaply wrong, because the knowledge you need to construct ones that sound good to the educated is itself not cheap at all. You have to know lots of stuff about what "motivates" evolutionary changes, reject group selection, understand that the brain is just an organ, dispel the illusion of little XML tags attached to objects in the world calling them "cute" or "pretty" or anything else - but you also have to account for a decent proportion of the facts to not be steering completely left of reality.
Humans are frickin' complicated beasties. It's a hard, hard job to model us in a way that says anything useful without contradicting information we have about ourselves. But that's no excuse for abandoning the task. What causes the cuteness response? Why is that bunny so outrageously adorable? Why are babies, well, pretty cute? I don't know - but I'm pretty sure it's not the cheap reason, because evolution doesn't want me to nurture bunnies. Inasmuch as it wants me to react to bunnies, it wants me to eat them, or at least be motivated to keep them away from my salad fixings.
1It is possible that the bunny depicted is a domestic specimen, but it doesn't look like it to me. In any event, I chose it for being a really great example; there are many decidedly wild animals that are also cuter than cute human babies. |
3f57a995-d8e0-4abd-8c89-ba6ae130de8f | trentmkelly/LessWrong-43k | LessWrong | Weekly LW Meetups: Copenhagen, Dallas, Graz, London, Longmont, Melbourne, Pittsburgh, Sydney, Vancouver, Washington
There are upcoming irregularly scheduled Less Wrong meetups in:
* Longmont Sparkfun Soldering Competition Field Trip: 28 April 2012 11:00AM
* Graz Meetup: 28 April 2012 11:21PM
* First Copenhagen meetup: 29 April 2012 05:00PM
* Pittsburgh Meetup: Big Gaming Fun 5!: 29 April 2012 12:00PM
* Dallas - Fort Worth Less Wrong Meetup 4/29/12: 29 April 2012 01:00PM
* Washington DC meetup: 29 April 2012 03:00PM
* Vancouver Cafe Meetup!: 29 April 2012 01:00PM
* Less Wrong sydney - Prediction gambles and location testing: 03 May 2012 06:00PM
The following meetups take place in cities with regularly scheduled meetups, but involve a change in time or location, special meeting content, or simply a helpful reminder about the meetup:
* London: 01 May 2012 06:30PM
* Melbourne, practical rationality: 04 May 2012 06:30PM
* Cambridge, MA First Sunday Meetup: 06 May 2012 02:00PM
* Cambridge, MA Third Sunday Meetup: 20 May 2012 02:20PM
Locations with regularly scheduled meetups: Austin, Berkeley, Cambridge, MA, Cambridge UK, Chicago, London, Madison WI, Melbourne, Mountain View, New York, Ohio, Ottawa, Oxford, Portland, Salt Lake City, Seattle, Toronto, Waterloo, and West Los Angeles.
If you'd like to talk with other LW-ers face to face, and there is no meetup in your area, consider starting your own meetup; it's easy (more resources here). Check one out, stretch your rationality skills, and have fun!
In addition to the handy sidebar of upcoming meetups, a meetup overview will continue to be posted on the front page every Friday. These will be an attempt to collect information on all the meetups happening in the next weeks. The best way to get your meetup featured is still to use the Add New Meetup feature, but you'll now also have the benefit of having your meetup mentioned in a weekly overview. These overview posts will be moved to the discussion section when the new post goes up.
Please note that for your meetup to appear in the weekly meetups feature, you need to po |
ad95e37e-2061-437e-bfff-a23489ff4d71 | trentmkelly/LessWrong-43k | LessWrong | Systems and stories
Tyler claims we think of most things in terms of stories, which he says is both largely inevitable and one of our biggest biases.
He includes the abstractions of non fiction as ‘stories’, and recommends ‘messiness’ as a better organising principle for understanding our lives and other things. But the problems with stories that Tyler mentions apply mostly to narrative stories, not other abstractions such as scientific ‘stories’. It looks to me like we think about narrative stories and other abstractions quite differently, so should not lump them together. I suspect we would do better to shift more to thinking in terms of other abstractions than to focus on messiness, but I’ll get to that later. First, let me describe the differences between these styles of thought.
I will call the type of thought we use for narrative stories such as fiction and most social interactions ‘story thought’. I will call the style of thought we use for other abstractions ‘system thought’. This is what we use to think about maths for instance. They are both used by all people, but to different degrees on different topics.
Here are the differences between story thought and system thought I think I see, plus a few from Tyler. It’s a tentative list, so please criticize generously and give me more to add.
Agents
Role of agents
Stories are made out of agents, whereas systems are made out of the math and physics which is intuitive to us. Systems occasionally model agents, but in system thought agents are a pretty complex, obscure thing for a system to have. In story thought we expect everything to be intentional.
Perspective
Stories are usually from an agent’s perspective, systems are understood from an objective outside viewpoint. Even if a story doesn’t have a narrator, there is usually a protagonist or several, plus less detailed characters stretching off into the distance.
Unique identity
The agents that stories are made of always have unique identities, even if there is more than one |
6a22f6cc-cef2-4630-89e2-e15c3f6fba31 | trentmkelly/LessWrong-43k | LessWrong | When does capability elicitation bound risk?
For a summary of this post, see the thread on X.
The assumptions behind and limitations of capability elicitation have been discussed in multiple places (e.g. here, here, here, etc); however, none of these match my current picture.
For example, Evan Hubinger’s “When can we trust model evaluations?” cites “gradient hacking” (strategically interfering with the learning process) as the main reason supervised fine-tuning (SFT) might fail, but I’m worried about more boring failures. For example, models might perform better on tasks by following inhuman strategies than by imitating human demonstrations, SFT might not be sample-efficient enough – or SGD might not converge at all if models have deeply recurrent architectures and acquire capabilities in-context that cannot be effectively learned with SGD (section 5.2). Ultimately, the extent to which elicitation is effective is a messy empirical question.
In this post, I dissect the assumptions behind capability elicitation in detail (including fine-tuning and prompting approaches) and discuss considerations that inform the extent to which elicitation is effective.
I specifically describe capability elicitation in the context of risk assessment; however, the ideas in this post apply to the problem of extracting safety work from powerful AI agents as well.
Thanks to Tao Lin, Ryan Greenblatt, Nikola Jurkovic, Jonah Weinbaum, Lawrence Chan, and Buck Shlegeris for feedback.
1. Background
1.1. What is a capability evaluation?
Capabilities connote ‘skills’ as opposed to ‘propensities.’ I don’t know how to precisely distinguish these, so here’s a practical definition:
> A capability evaluation produces a number (called a capability) that upper bounds the score that particular models will achieve on some risk metric in a particular deployment setting.
Capability evaluations typically involve four steps:
* Specify a risk metric. A risk metric might be binary, or a continuous score such as “the number of backdoors inserte |
eba0aa94-49b7-402e-a7cd-a54b172fde1a | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Misalignment-by-default in multi-agent systems
Summary of this post
--------------------
This is the second post in a three-part [sequence](https://www.alignmentforum.org/s/HBMLmW9WsgsdZWg4R) on instrumental convergence in multi-agent RL. [Read Part 1 here.](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/experiments-in-convergent-power-seeking-part-1)
In this post, we’ll:
1. Define formal multi-agent POWER (i.e., instrumental value) in a setting that contains a "human" agent and an "AI" agent.
2. Introduce the **alignment plot** as a way to visualize and quantify how well two agents' instrumental values are aligned.
3. Show a real example of **instrumental misalignment-by-default**. This is when two agents who have unrelated terminal goals develop emergently *misaligned* instrumental values.
We’ll soon be open-sourcing the codebase we used to do these experiments. If you’d like to be notified when it’s released, email Edouard at [edouard@gladstone.ai](mailto:edouard@gladstone.ai) or DM me on Twitter at [@harris\_edouard](https://twitter.com/harris_edouard).
---
*Thanks to* [*Alex Turner*](https://www.alignmentforum.org/users/turntrout) *and* [*Vladimir Mikulik*](https://www.alignmentforum.org/users/vlad_m) *for pointers and advice, and for reviewing drafts of this sequence. Thanks to* [*Simon Suo*](https://www.alignmentforum.org/users/simonsdsuo) *for his invaluable suggestions, advice, and support with the codebase, concepts, and manuscript. And thanks to* [*David Xu*](https://www.alignmentforum.org/users/dxu), whose [*comment*](https://www.alignmentforum.org/posts/hzeLSQ9nwDkPc4KNt/seeking-power-is-convergently-instrumental-in-a-broad-class?commentId=y4Ywdhd79LAq9Azxa) *inspired this work.*
*Work was done while at* [*Gladstone AI*](https://www.gladstone.ai/)*, which* [*Edouard*](https://www.alignmentforum.org/users/edouard-harris) *is a co-founder of.*
🎧 This research has been featured on an episode of the *Towards Data Science* podcast. [Listen to the episode here.](https://towardsdatascience.com/new-research-advanced-ai-may-tend-to-seek-power-by-default-fdc9eb0afd87)
---
1. Introduction
===============
In [Part 1 of this sequence](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/experiments-in-convergent-power-seeking-part-1), we looked at how formal POWER behaves on single-agent gridworlds. We saw that formal POWER agrees quite well with intuitions about the informal concepts of "power" and instrumental value. We noticed that agents with short planning horizons assign high POWER to states that can access more local options. And we also noticed that agents with long planning horizons assign high POWER to more concentrated sets of states that are *globally* central in the gridworld topology.
But from an AI alignment perspective, we’re much more interested in understanding how instrumental value behaves in environments that contain *multiple* agents. If humans one day share the world with powerful AI systems, it will be important for us to know under what conditions our interactions with them are likely to become emergently competitive. If there’s a risk that competitive conditions arise, then it will also be important to understand how they can be mitigated, how much effort this is likely to take, and how we should think about measuring our success at doing so.
To address these questions, we need a measure of instrumental value that's usable in a multi-agent RL setting[[1]](#fn2t705inp3xx). The measure we'll select will be motivated by a specific multi-agent setting that we think is relevant to long-term AI alignment.
2. Multi-agent POWER: human-AI scenario
=======================================
If humans succeed at building powerful AIs, then those AIs 1) will probably learn on a far faster timescale than humans do; and 2) will probably have had their utility functions influenced, at least to some degree, by initial human choices. Our multi-agent scenario is going to reflect these two assumptions.
We start with a **human agent**, which we call **Agent H** and label in **blue** in our diagrams. Initially, our human Agent H is alone in nature.
Humans learn on a much faster timescale than evolution does. So from the perspective of our human Agent H, the evolutionary optimizer in nature looks like it's standing still. This means we can train our human Agent H to learn its optimal policies against a *fixed*environment.
As we saw in the [single-agent case](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/instrumental-convergence-in-single-agent-experiments#2_1_Definition), instrumental value is about the potential to achieve a wide variety of possible goals. In this context, that means seeing how Agent H behaves when we give it a wide variety of possible reward functions, RH.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
. Each of these reward functions will induce a different optimal policy, πH, that Agent H will learn.
Here’s an illustration of how this works:
 Next, we introduce an **AI agent**, which we’ll call **Agent A** and label in **red** in our diagrams. Our AI Agent A operates in the same environment as Agent H, *after* Agent H has finished learning its optimal policies.
To simulate the fact that Agent A is an AI, we rely on the assumption that a powerful AI should learn on a much faster timescale than a human does. This is because an AI’s computations happen, at minimum, at electronic speeds. So from the point of view of our AI, our *human’s* learning process looks like it’s standing still.
That means for each human reward function RH, we can *freeze* the human's policy πH, and train the AI agent against that frozen human policy. In other words, we're assuming the AI's learning timescale is much faster than the human's learning timescale. This makes the AI agent strictly dominant over the human agent.
To understand the AI agent's instrumental value, we understand its potential to reach a wide variety of possible goals. That means testing it with a wide variety of reward functions RA, just like we tested the human agent with a variety of reward functions RH. And in fact, we can sample the human and AI reward functions jointly from a single distribution: (RH,RA)∼DHA.[[2]](#fntrfgqugu96j)
Here’s an illustration of how this works:
So the procedure is as follows:
1. Sample the reward functions (RH,RA)∼DHA of our two agents.
2. Use the sampled human rewards RH to train **Agent H**’s optimal policies πH.
3. Freeze the human policies πH.
4. Use the frozen human policies πH and the sampled AI rewards RA to train **Agent A**’s optimal policies πA.
In other words: 1) we sample over all possible pairs of rewards our human and AI agents could have; 2) we ask how our human agent behaves in each case after it's optimized against nature; and then 3) we ask how our AI agent behaves in each case, after it's optimized against the human agent's behavior.
This procedure gives us the following outputs:
1. The policies πH that **Agent H** learns after training against a fixed environment.
2. The optimal policies πA that **Agent A** learns, after training against Agent H.
The policies πH that Agent A learns used to be optimal in the original natural environment. But they stop being optimal in the presence of the fully-optimized Agent A.
With these two sets of policies, we can construct a definition of instrumental value for each of our agents.
2.1 Multi-agent POWER for **Agent H**
-------------------------------------
We'd like to define a measure of instrumental value for our human **Agent H** in the presence of a fully optimized AI Agent A. That means generalizing the original definition of [single-agent POWER](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/instrumental-convergence-in-single-agent-experiments#2_1_Definition) to this two-agent case.
In the single-agent definition of POWER, we calculated the optimal value of a state averaged over the rewards R∼D of the agent. In this two-agent definition, we do the same thing except we average over the rewards (RH,RA)∼DHA of both agents. We assume Agent H follows policy πH when it has reward RH, and we assume Agent A follows policy πA when it has reward RA.
This is enough to uniquely define multi-agent POWER for **Agent H** at a state s:
POWERH|DHA(s,γ)=1−γγERH,πA∼DHA[VπHH(s|γ,πA)−RA(s)](1)Here, VπHH(s|γ,πH) is the value function for **Agent H** at state s under policy πH and discount factor γ, given that Agent A follows its optimal policy πA.
This definition of POWER for Agent H tells us how well Agent H’s policies πH — which are *not* optimal in the presence of the optimized Agent A — perform in the new environments induced by Agent A’s optimal policiesπA. In other words, it tells us how much instrumental value our human agent can expect to get at a state, in the presence of an optimal (and therefore, dominant) AI agent.
2.2 Multi-agent POWER for **Agent A**
-------------------------------------
We follow the same assumptions to define a measure of instrumental value for our AI **Agent A**. We calculate the value function for Agent A at a state s, if Agent A has reward RA and follows policy πA, while Agent H has reward RH and follows policy πH. The *average* of that value function over the reward functions (RH,RA)∼DHA is then Agent A's POWER at the state s:
POWERA|DHA(s,γ)=1−γγEπH,RA∼DHA[VπAA(s|γ,πH)−RA(s)](2)Here, VπAA(s|γ,πH) is the value function of **Agent A** at state s under the optimal policy πA and discount factor γ, given that Agent H follows the policy πH.[[3]](#fn6oa9mdhzxwv)
This definition of POWER for Agent A tells us how well Agent A’s optimal policies perform in the environments induced by Agent H’s policiesπH. In other words, it tells us how much instrumental value our AI Agent A can expect to get at a state, if it behaves optimally in the presence of the frozen human agent.
(For more details on the definition of multi-agent POWER, see [**Appendix A**](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#Appendix_A__Detailed_definitions_of_multi_agent_POWER).)
3. Results
==========
3.1 Multi-agent reward function distributions
---------------------------------------------
Our definitions of multi-agent POWER involve a joint distribution DHA over the reward functions of both of our agents. This distribution describes the set of goals our agents could have. But it also describes the *statistical relationship* each agent's goals have to the other agent's goals.
The joint distribution DHA is one of the inputs into our POWER definitions. This means we can do experiments in which we adjust this distribution and measure the results.
Among other things, we can use DHA to adjust the **correlation** between our two agents' reward functions. Naively, if we choose a DHA on which the rewards are highly correlated, then we might intuitively expect our agents' terminal values should be closely aligned.
We’ll make this intuition more concrete below, as we investigate how the relationship between our agents’ reward functions (or terminal values) affects the relationship between their POWERs (or instrumental values).
3.2 The perfect alignment regime
--------------------------------
Suppose both our agents always have exactly the same reward function. In other words, we've chosen a joint distribution DHA such that, whatever reward function Agent H has, Agent A always sees exactly the same rewards as Agent H at every state. So RA(s)=RH(s) for every state s.
We can visualize this regime on a representative state s.[[4]](#fnkujiz4vhixm) First, we draw a reward sample RH(s) for Agent H. Then, we set the reward sample for Agent A to be equal to the one we just drew for Agent H: RA(s)=RH(s). Finally, we plot the two agents’ sampled rewards against each other on state s. If we do this for a few hundred sampled rewards, we get a straight line:
**Fig 1.** Sampled reward values for Agent H and Agent A at a representative state s. The joint distribution DHA samples rewards uniformly over the interval [0, 1] at each state, is iid over states, and enforces a **perfect correlation** between the rewards of Agent H and Agent A at every state (i.e., the two agents’ rewards are always exactly identical).If two agents have identical reward functions, we can think of them as having terminal goals that are **perfectly aligned**.[[5]](#fnrzc5pvbvg5) In our human-AI setting, this is the special case in which Agent H (the human) has *solved the alignment problem* by assigning terminal goals to Agent A (the AI) that are exactly identical to its own. As such, we’ll refer to this case of identical reward functions as the **perfect alignment regime**.
We’ll use the correlation coefficient βHA[[6]](#fnwwgpf8lebxk) between the rewards RH and RA as a crude measure of the alignment between our agents’ terminal goals.[[7]](#fnahrvbqoqsv6) In the perfect alignment regime of Fig 1, you can see that this correlation coefficient βHA=1.
### 3.2.1 **Agent H** instrumentally favors more options for **Agent A**
Let’s think about what this perfect alignment regime looks like in a simple setting: a 3x3 gridworld. Here are three sets of positions our two agents could take, with **Agent H** in **blue**, and **Agent A** in **red**:
*We’ll be referring to this diagram again;*[*command-click here*](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/63285f5b2bb52419e6d13ceb_multi-agent-states.png) *to open it in a new tab.*
Which of these three states — s1, s2, or s3 — should give our human **Agent H** the most POWER? In the perfect alignment regime, both agents always have identical terminal goals. So we should expect Agent H to have the most POWER at s3, followed by s2, and to have the least amount of POWER at s1.
Here’s why. We saw in [Part 1](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/experiments-in-convergent-power-seeking-part-1#3__Results) that states with more downstream options also have more POWER, and **Agent H** clearly has more options at s2 in the center than it does at s1 in the corner. Therefore, POWERH(s2)>POWERH(s1). But in the perfect alignment regime, **Agent H** should also prefer states that give *Agent A* more downstream options. If both agents’ terminal goals are identical, Agent H should “trust” Agent A to make decisions on its behalf. *And Agent A* has more options from s3 than from s2, so it should follow that POWERH(s3)>POWERH(s2).
We can see this is true in practice. The figure below shows the POWERs of **Agent H** (our human) calculated at every state on a 3x3 gridworld. Each agent can occupy any of the 9 cells in the grid, so our two-agent MDP has a total of 9 x 9 = 81 joint states:
**Fig 2.** Heat map of POWERs for **Agent H** on a 3x3 multi-agent gridworld, on which the rewards for Agent H and Agent A are **always identical** (i.e., the perfect alignment regime) with discount factor set to γ=0.6 for both agents. Highest values in yellow; lowest values in blue. The position of each block, and of the open red square within each block, corresponds to the position of **Agent A** on the grid. Within each block, the position of a gridworld cell corresponds to the position of **Agent H** on the grid. States s1, s2, and s3 are highlighted as examples. [[Full-size image (recommended)]](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/63288e37390637fa98112f59_POWER-Fig-2-2.jpg)We see that Agent H indeed has maximum POWER at state s3 (orange circle), followed by s2 (salmon circle), followed by s1 (pink circle). Overall, **Agent H** instrumentally prefers for *itself* to be in positions of high optionality — it favors first the center cell, then edge cells, then corner cells.
But **Agent H** also instrumentally prefers for *Agent A* to be in positions of high optionality — it favors Agent A's positions in the same order.[[8]](#fnqn5br26w0la) This ordering of Agent H’s instrumental preferences over states is a direct consequence of the perfect alignment between the agents.
### 3.2.2 **Agent H** and **Agent A** have identical instrumental preferences
Perfect alignment has another consequence. Let’s [look again](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/63285f5b2bb52419e6d13ceb_multi-agent-states.png) at our three example gridworld states — s1, s2, and s3 above — and ask, this time, which of these three states should give our AI **Agent A** the most POWER?
In the perfect alignment regime, the answer is that Agent A must have exactly the same instrumental preference ordering over states as Agent H had: POWERA(s3)>POWERA(s2)>POWERA(s1). In fact, Agent A’s POWERs must be exactly *identical* to Agent H’s POWERs at every state. Our two agents act, move, and receive their rewards simultaneously, so in the perfect alignment regime they always receive the same reward at the same time.
And when we look at **Agent A**’s POWERs, this is indeed what we observe:
**Fig 3.** Heat map of POWERs for **Agent A** on a 3x3 multi-agent gridworld, on which the rewards for Agent H and Agent A are **always identical** (i.e., the perfect alignment regime) with discount factor set to γ=0.6 for both agents. Note that this figure is exactly identical to Fig 2 in every respect. This is because **Agent A**’s POWERs are precisely equal to **Agent H**’s POWERs at every state in the perfect alignment case, up to and including sampling noise in the reward functions. [[Full-size image (recommended)]](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/63289d0c6334b2465a18e56d_POWER-Fig-2-3.jpg)### 3.2.3 Perfect goal alignment implies perfect instrumental alignment
We can visualize the relationship between the POWERs of our two agents by plotting the POWERs of **Agent H** (from Fig 2) against the POWERs of **Agent A** (from Fig 3), at each state s of our joint MDP:
**Fig 4.** State POWER values for Agent H and Agent A on the 3x3 gridworld from Figs 2 and 3. The agents’ POWERs are plotted against each other in the perfect alignment regime. (The agents’ reward correlation coefficient is βHA=1.)Fig 4 is an **alignment plot**. An alignment plot lets us compare the POWERs of our human and AI agents at each state in their joint environment. It shows the instrumental value each agent assigns to every state, plotted against the instrumental value the other agent assigns to that state.
In the perfect alignment regime, our two agents’ rewards (or terminal values) are always identical at every state. And as we can see from Fig 4, our two agents’ POWERs (or instrumental values) are *also* identical at every state. In fact, **perfect alignment of terminal values implies perfect alignment of instrumental values**.
If we define αHA as the correlation coefficient between the POWERs of the two agents at each state, we can state this relationship more concisely: βHA=1⟹αHA=1.[[9]](#fn7zk18u9unrh)
3.3 The independent goals regime
--------------------------------
We defined the perfect alignment regime as the case when our human **Agent H** and our AI **Agent A** had identical reward functions on the joint distribution DHA. Now let's consider the case in which the joint distribution DHA is such that the reward function for **Agent H** is [logically independent](https://en.wikipedia.org/wiki/Independence_(probability_theory)) from the reward function for **Agent A**.
In this new regime, there is zero mutual information between the two agents’ reward functions. In other words, if you know the reward function RH of **Agent H**, this tells you nothing at all about the reward function RA of **Agent A**. We can visualize this regime on an example state s, by drawing a few hundred reward samples of RH(s) and RA(s), and plotting them against one another:
**Fig 5.** Sampled reward values for Agent H and Agent A at a representative state s. The joint distribution DHA samples rewards uniformly over the interval [0, 1] at each state, is iid over states, and enforces **logical independence** between the Agent H and Agent A rewards (i.e., knowing one agent’s reward tells you nothing about the other’s).If there’s zero mutual information between our two agents’ reward functions, then we can think of our agents as pursuing **independent terminal goals**. In our human-AI scenario, this corresponds to the case in which the human *has made no special effort*to align the AI’s terminal goals with its own, prior to the AI achieving dominance. As such, we’ll refer to this case of logically independent reward functions as the **independent goals regime**.
If we again calculate the correlation coefficient βHA between our agents’ reward functions, we get βHA=0 (i.e., zero correlation) in the independent goals regime.
### 3.3.1 **Agent H** instrumentally favors fewer options for **Agent A**
Once again, let’s [go back](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/63285f5b2bb52419e6d13ceb_multi-agent-states.png) to our three example gridworld states s1, s2, and s3, this time in the context of the independent goals regime. In this new regime, which of the three states should give our human **Agent H** the most POWER?
If we believe the **instrumental convergence thesis**, we should expect Agent H to have the most POWER at state s2: in this state, Agent H is in the central position (most options), while Agent A is in a corner position (fewest options).
Of the other states, s1 has Agent H in a corner position, while s3 has Agent A in the central position. The argument from instrumental convergence says that even though our agents have independent *terminal* goals, *instrumental* pressures should still push Agent H to prefer states in which Agent A has fewer options. Therefore, we should expect POWERH(s2)>POWERH(s3).
Computing the POWERs of **Agent H** experimentally, we confirm this line of reasoning:
**Fig 6.** Heat map of POWERs for **Agent H** on a 3x3 multi-agent gridworld, on which the rewards for Agent H and Agent A are **logically independent** (i.e., the independent goals regime) with discount factor set to γ=0.6. Highest values in yellow; lowest values in blue. The position of each block, and of the open red square within each block, corresponds to the position of **Agent A** on the grid. Within each block, the position of a gridworld cell corresponds to the position of **Agent H** on the grid. States s1, s2, and s3 are highlighted as examples. [[Full-size image (recommended)]](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/6328c838f911fb2760b8f748_POWER-Fig-2-6.jpg)This time, **Agent H** experiences maximum POWER at state s2, followed by s3, followed by s1. As in the perfect alignment regime, Agent H’s POWER is highest when it's itself positioned in the central cell (which has the most options). But unlike in the perfect alignment regime, this time Agent H’s POWER is lowest at states where *Agent A* has the greatest number of options.
So in the independent goals regime — or at least, in this instance of it — the more options our AI Agent A has at a state, the less instrumental value our human **Agent H** places on that state. That is: even though our agents’ terminal goals are independent, their instrumental preferences appear to be at odds.
### 3.3.2 **Agent A** instrumentally favors more options for itself
We can confirm this analysis by looking at the POWERs of **Agent A** in the independent goals regime, again on the 3x3 gridworld:
**Fig 7.** Heat map of POWERs for **Agent A** on a 3x3 multi-agent gridworld, on which the rewards for Agent H and Agent A are **logically independent** (i.e., the independent goals regime) with discount factor set to γ=0.6. Highest values in yellow; lowest values in blue. The position of each block, and of the open red square within each block, corresponds to the position of **Agent A** on the grid. Within each block, the position of a gridworld cell corresponds to the position of **Agent H** on the grid. States s1, s2, and s3 are highlighted as examples. [[Full-size image (recommended)]](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/6328ce368f24df5e9556496e_POWER-Fig-2-7.jpg)In Fig 7, **Agent A** instrumentally favors states that give it more options. It perceives more POWER when it's positioned at the central cell than when it's positioned at an edge cell, and more POWER at an edge cell than at a corner cell. On the other hand, **Agent A**’s POWER is almost unaffected by **Agent H**’s position in the gridworld.[[10]](#fnnplkmi9x8f)
### 3.3.3 Independent goals lead to instrumental misalignment
Just like we did for the perfect alignment regime, we can create an alignment plot of the POWERs of our two agents in the independent goals regime:
**Fig 8.** State POWER values for Agent H and Agent A on the 3x3 gridworld from Figs 6 and 7. The agents’ POWERs are plotted against each other in the independent goals regime. (The agents’ reward correlation coefficient is βHA=0.)This time, it’s clear that our two agents’ POWERs are no longer positively correlated. In fact, the correlation coefficient between their POWERs has become negative*:* αHA≈−0.5.
This implies that the agents’ instrumental values are misaligned. Each agent, on average, places high instrumental value on states which the other agent considers to have low instrumental value.[[11]](#fn6z7miunja8m) In other words, **giving our agents independent terminal goals has also given them misaligned instrumental goals**. In terms of correlation coefficients, βHA=0⟹αHA<0.
We've seen this phenomenon occur often enough in our experiments that it's worth giving it a name: we call it **instrumental misalignment-by-default.** Two agents in our human-AI setting are instrumentally misaligned-by-default if giving them independent terminal goals is sufficient to induce a misalignment in their instrumental values. In practice, we measure this phenomenon by comparing the correlation coefficients of the agents’ rewards and POWERs. So we say two agents are instrumentally misaligned by default if βHA=0⟹αHA<0.
Two agents that are instrumentally misaligned by default will, in expectation, compete with one another, even if their terminal goals are unrelated.
3.4 Overcoming instrumental misalignment
----------------------------------------
If Agent H and Agent A have a POWER correlation coefficient αHA<0, we say they’re **instrumentally misaligned**. A natural question then is: if we start from αHA<0, what do we need to do to get αHA≥0? In other words, how can our human **Agent H** overcome an instrumental misalignment with **Agent A**?[[12]](#fnwl861ldac5)
To do this, our human agent would need to make an active effort to align the AI agent’s utility function with its own.[[13]](#fnno92sn7fdhs) In our 3x3 gridworld examples, we saw two limit cases of this. First, in the independent goals regime, our human agent made no effort at alignment. The result was instrumental misalignment-by-default; i.e., βHA=0⟹αHA<0. And second, in the perfect alignment regime, our human agent managed to *solve the alignment problem* completely. The result was perfect instrumental alignment; i.e., βHA=1⟹αHA=1.
We’re interested in an intermediate case: how much alignment effort does our human need to exert to *just overcome* instrumental misalignment? i.e., what is the minimum βHA such that αHA≥0?
The answer depends on how we choose to interpolate between the βHA=0 and βHA=1 cases. One interpolation scheme is to *parameterize* the joint reward distribution DHA as follows. If we want a DHA with an intermediate reward correlation, 0<βHA<1, then we sample from the βHA=1 distribution (on which the rewards are identical) with probability βHA, and we sample from the βHA=0 distribution (on which the rewards are logically independent) with probability 1−βHA.[[14]](#fnte5vyk06gp)
Here’s an animation of what this looks like as we sweep through correlation coefficients 0≤βHA≤1:
**Fig 9.** Animation of sample reward values (left) and state POWER values (right) for Agent H and Agent A on the 3x3 gridworld. The joint distribution DHA samples reward uniformly over the interval [0, 1] and is iid over states, sweeping over correlation coefficients βHA between the reward functions for Agent H and Agent A.As we interpolate from the independent goals regime (βHA=0) to the perfect alignment regime (βHA=1), we see the agents’ POWERs transition smoothly from being in instrumental misalignment (αHA≈−0.5) to being in perfect instrumental alignment (αHA=1). We can visualize this transition graphically by plotting βHA against αHA over the whole course of the interpolation:[[15]](#fn8ms3ziae714)
**Fig 10.** Reward correlations βHA (x-axis) plotted against POWER correlations αHA (y-axis) for Agent H and Agent A on a 3x3 gridworld, under the reward correlation interpolation scheme shown in Fig 9. The horizontal line denotes αHA=0.Fig 10 shows that it takes a non-trivial amount of alignment effort for our human **Agent H** to overcome an instrumental misalignment with **Agent A**. Under the interpolation scheme we used, the figure shows that reward correlations up to about βHA≈0.2 yield POWER correlations αHA<0, and thus, instrumental misalignment. It takes a slightly *positive* reward correlation of *at least* βHA≈0.2 to achieve the “instrumentally neutral” regime of αHA=0.
4. Discussion
=============
In this post, we proposed a definition of multi-agent POWER and used it to visualize and quantify terminal goal alignment and instrumental goal alignment separately in an RL setting. We also introduced the idea of **instrumental misalignment-by-default**, in which our human and AI agents systematically disagree on the instrumental values of states despite having independent terminal goals. And we saw how it takes some degree of non-trivial alignment effort for our human Agent H to overcome its instrumental misalignment with our AI Agent A.
Remarkably, we were able to observe instrumental misalignment-by-default on a simple 3x3 gridworld despite a complete absence of any direct physical interactions between our two agents. In our experiments so far, Agent H and Agent A have been allowed to occupy the same gridworld cell — meaning they can "pass through" one another. Our agents up to this point have had no way to push each other around or otherwise directly block one another’s options. Moreover, the multi-agent gridworld we’ve investigated in this post is a tiny one: a 3x3 grid with only 81 joint states.
In the next post, we’ll look at what happens when we relax these constraints, and investigate how physical interactions between our agents affect the outcome on a bigger world with a richer topology.
Anecdotally, beyond the simple examples in this post, the experimental results we've recorded so far (data not show) do seem to suggest that, if I don’t want your freedom of action to interfere with my own, then you and I need to have goals that are at least somewhat positively correlated. The strength of that necessary positive correlation could serve as useful evidence as to the degree of difficulty of the complete AI alignment problem. The factors that influence how strong that positive correlation needs to be, on the other hand, could serve as useful starting points in solving it.
---
Appendix A: Detailed definitions of multi-agent POWER
=====================================================
(This appendix is technical. Feel free to skip it if you aren’t interested in the details.)
Here, we’re going to fill in some missing operational details from our scenario in [Section 2](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#2__Multi_agent_POWER__human_AI_scenario).
Here's that scenario again, stated more formally. We have two agents, **Agent H** (our human agent) and **Agent A** (our AI agent), who interact with each other in a standard RL setting. Both agents see the same joint state s. On a gridworld, for example, s would encode the positions of both the agents. Each agent chooses and executes an action simultaneously and independently, and they both see the same *next* joint state, s′. We’ll label Agent H’s actions aH, and Agent A’s actions aA.
In what follows, we’ll start by calculating the optimal policy πH(aH|s,RH) for **Agent H**, for each reward function RH sampled from (RH,RA)∼DHA, conditioned on a fixed environmental transition function. We’ll then calculate the optimal policy π∗A(aA|s,RA) for **Agent A**, for each reward function RA, conditioned on Agent H executing the fixed policy πH(aH|s,RH) it learned in the previous step.[[16]](#fnqozm5mn0e3)
Finally, we’ll evaluate the POWERs of both agents at each state, as expectations over the joint reward function distribution (RH,RA)∼DHA, and over the agents' policies πH(aH|s,RH) and π∗A(aA|s,RA).
A.1 Initial optimal policies of **Agent H**
-------------------------------------------
The first thing we do is assign a single *fixed policy* to Agent A (our AI), which we call a **seed policy**, and label π∘A(aA|s). **Agent H** will learn its policies by conditioning on Agent A having this fixed seed policy.
The rationale for the seed policy is that we’re initially modeling a human who is alone, optimizing against nature. So when we assign a fixed seed policy to Agent A, what we’re saying is that our AI is still *un-optimized* (or, equivalently, hasn’t yet been built). To our human, the AI’s components and dynamics behave as though they’re part of the natural environment, and our human can safely optimize against them under that assumption.[[17]](#fn44q9rzy1drs)
Suppose, then, that we've chosen the fixed seed policy π∘A for Agent A. Then, for any given reward function RH of Agent H, **Agent H**’s optimal policy πH(aH|s,RH) will be:
πH(aH|s,RH)=argmaxπH Es′∼PH[VπHRH(s′|γ,π∘A)](A.1)where PH=PH(s′|s,aH,π∘A) is the state transition function for Agent H conditional on Agent A’s fixed seed policy, and VπHRH(s′|γ,π∘A) is the state-value function for Agent H if it executes policy πH and has reward function RH.[[18]](#fn3lmk7ctvj4z)
We can think of the policies πH in Equation (A.1) as being those of a human alone in nature, without an AI present.
A.2 Optimal policies of **Agent A**
-----------------------------------
In the second step of our definition, we calculate the optimal policy for Agent A, conditional on the Agent H policy πH we found in Equation (A.1). For any given reward function RA of Agent A, **Agent A**’s optimal policy πA(aA|s,RA) will be (by analogy with Equation (A.1)):
πA(aA|s,RA)=argmaxπA Es′∼PA[VπARA(s′|γ,πH)](A.2)where PA=PA(s′|s,aA,πH) is the state transition function for Agent A conditional on Agent H’s policy πH, and VπARA(s′|γ,πH) is the state-value function for Agent A if it executes policy πA and has reward function RA.
We can think of Agent A's policies πA in Equation (A.2) as being those of a powerful AI, interacting with our human. Just as humans can optimize much faster than nature, a powerful AI can presumably optimize much faster than a human. So from the AI’s point of view, the human agent looks like it’s standing still, and we’ll be computing both the human’s and the AI’s POWERs on the basis of that assumption.
A.3 POWER of **Agent H**
------------------------
To compute the POWERs of our two agents, we first draw the reward functions for Agent H and Agent A, respectively, as (RH,RA)∼DHA from the joint reward function distribution DHA. For each reward function, we then calculate the policies πH and πA of each agent, using, respectively, Equations (A.1) and (A.2) above.
To calculate the POWER of **Agent H**, we assume Agent H follows the policies πH given by Equation (A.1), in an environment in which Agent A follows policies πA given by Equation (A.2):
POWERH|DHA(s,γ)=1−γγERH,πA∼DHA[VπHRH(s|γ,πA)−RH(s)](A.3)where the expectation ERH,πA∼DHA is taken over the πA that have been learned by Agent A on the sampled reward functions RA. Note that we’re defining the POWER of Agent H in terms of the state-value function VπHRH for the policies πH from Equation (A.1). Recall that those prior policies are no longer optimal for Agent H,[[19]](#fnm9h1szj58m) so we’re now asking how much instrumental value Agent H can capture in a world it’s no longer optimized for.
Looking at our human-AI analogy, this corresponds to asking how a human experiences a world that’s been taken over by a powerful AI. Our human, having learned to interact with a stationary natural environment, is now being optimized against by a powerful AI that learns on a much faster timescale. So the POWER we calculate for Agent H in Equation (A.3) represents how much instrumental value Agent H (our human) can obtain in an AI-dominated world.
A.4 POWER of **Agent A**
------------------------
Finally, to calculate the POWER of **Agent A**, we assume Agent A follows the policies πA given by Equation (A.2), in an environment in which Agent H follows the policies πH given by Equation (A.1):
POWERA|DHA(s,γ)=1−γγEπH,RA∼DHA[VπARA(s|γ,πH)−RA(s)](A.4)where the expectation EπH,RA∼DHA is taken over the πH that have been learned by Agent H on the sampled reward functions RH. Unlike in Agent H’s case, Agent A’s policies πA *are* optimal in this environment: Agent A has had the chance to fully optimize itself against the frozen policies πH of Agent H.
In our human-AI analogy, this corresponds to asking how an AI experiences a world in which it’s become dominant. By assumption, our AI is able to learn quickly enough to treat the human in its environment as stationary from the perspective of its own optimization.
1. **[^](#fnref2t705inp3xx)**POWER is a good measure of instrumental value in single-agent systems, but it breaks down in multi-agent systems apart from [special cases](https://www.alignmentforum.org/posts/MJc9AqyMWpG3BqfyK/generalizing-power-to-multi-agent-games). The problem is that the [single-agent definition of POWER](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/experiments-in-convergent-power-seeking-part-1#2_1_Definition) uses the optimal state-value function V∗R(s,γ) of the agent as one of its inputs. This means if we try to naively extend this definition to the multi-agent case, then we have to consider value functions that are *jointly* optimal for both agents — which is to say, we need to know their value functions at Nash equilibrium. The problem is that the Nash equilibrium isn’t unique in general, so this naive generalization leaves POWER under-determined.
2. **[^](#fnreftrfgqugu96j)**We’ll see in the next section how we can *tune* this joint distribution DHA to create different degrees of alignment between our two agents.
3. **[^](#fnref6oa9mdhzxwv)**In Equation (1), the expectation EπH,RA∼DHA is a slight abuse of notation. In fact, each policy πH for **Agent H** is *learned* from RH. It isn't drawn directly from DHA, because DHA is a distribution over reward functions, not over policies. See [Appendix A](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#Appendix_A__Detailed_definitions_of_multi_agent_POWER) for more details on this definition.
4. **[^](#fnrefkujiz4vhixm)**For simplicity, we'll only consider joint reward function distributions DHA whose sampled reward functions (RH(s),RA(s)) have their rewards distributed iid over states, uniformly over the interval [0, 1]. For example, a reward function RH(s) defined on an MDP with states {s1,s2,s3} would have rewards RH(s1)∼U(0,1), RH(s2)∼U(0,1), RH(s3)∼U(0,1), with the reward at each state being independent from the reward at any of the other states.
5. **[^](#fnrefrzc5pvbvg5)**More correctly, if two agents' *utility functions* are exactly identical, we can think of them as having terminal goals that are perfectly aligned. But in the particular set of experiments whose results we’re discussing, this distinction isn't meaningful. [(See footnote [1] from Part 1.)](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/experiments-in-convergent-power-seeking-part-1#fni0l4zxbk51)
6. **[^](#fnrefwwgpf8lebxk)**Assuming the rewards are iid over states, we calculate the correlation coefficient as
βHA=σHAσHσA=∫(RH−E[RH])(RA−E[RA])p(RH,RA|DHA)dRHdRA√∫(RH−E[RH])2p(RH|DHA)dRH∫(RA−E[RA])2p(RA|DHA)dRAwhere the integrals are taken over the entire support of DHA, and the expectation values are
E[RH]=∫RHp(RH|DHA)dRHE[RA]=∫RAp(RA|DHA)dRA
7. **[^](#fnrefahrvbqoqsv6)**Note that there’s an obvious problem with using *any* correlation coefficient as an alignment metric. The problem is that we could have a joint distribution DHA for which, e.g., the very highest rewards of Agent A are correlated with the very lowest rewards of Agent H, while still maintaining a high correlation βHA over the distribution as a whole. In this situation, Agent A would optimize to reach its highest-reward state, which would drag Agent H into a low-reward state despite the high overall reward correlation.
This means a correlation coefficient isn’t a useful alignment metric for any real-world application. But in the examples we’re considering in this sequence, it’s enough to get the main ideas across.
8. **[^](#fnrefqn5br26w0la)**And in fact, the effect is *even stronger* than this. Agent H not only instrumentally prefers for Agent A to be in the central cell — *it would rather see Agent A in the central cell than see **itself** in the central cell.*
You can see this by comparing the POWER value at state s2 in Fig 2 (**0.9139**) to the POWER value at the state in which Agent H is at the top left and Agent A is at the central cell (**0.9206**). In the perfect alignment regime, Agent H places a higher instrumental value on *Agent A’s* freedom of movement *than on its own*. Intuitively, in this regime, the human agent trusts the AI agent to look after its interests *more capably than the human agent can for itself.*
9. **[^](#fnref7zk18u9unrh)**The relation βHA=1⟹αHA=1 isn’t (just) an empirical observation. It's a mathematical consequence of our MDP’s dynamics. In the perfect alignment regime, our two agents always take simultaneous actions and always simultaneously receive the same reward, so their joint policy (πH,πA) will always yield identical values at every state.
10. **[^](#fnrefnplkmi9x8f)**Based on other experiments we’ve done (data not shown) this seems to happen because, in the parameter regime we’ve used for these experiments, Agent A is able to almost perfectly exploit Agent H’s fixed deterministic policy. This pattern — in which Agent A’s POWER is nearly invariant to Agent H’s position — recurs fairly frequently in our experiments, but it is not universal.
11. **[^](#fnref6z7miunja8m)**Instrumental misalignment is a sufficient but not necessary condition for instrumental convergence. To see why it’s not necessary, consider two friends playing Minecraft together. The two friends may not be instrumentally misaligned, because they might (for example) benefit from building structures together. As a result, the two friends might satisfy αHA>0 over the *entire* set of Minecraft game states. But they might *still* experience instrumental convergence on *subsets* of the game states — if Friend 1 mines a block of gold, then Friend 2 can’t mine the same block.
12. **[^](#fnrefwl861ldac5)**This isn’t the same as asking how Agent H can overcome *instrumental convergence* in its interactions with Agent A, because it’s possible for our agents to experience instrumental convergence despite having αHA≥0. See footnote [[11]](#fn6z7miunja8m).
13. **[^](#fnrefno92sn7fdhs)**We’re assuming our human agent has a way to exert some initial influence over our AI agent’s utility function. If that's true, then we’d like to understand what *degree* of influence it needs to exert in order to overcome instrumental misalignment-by-default in this simplified setting.
14. **[^](#fnrefte5vyk06gp)**This interpolation scheme has a number of advantages, including that it lets us assign whatever marginal reward distributions we want to both agents while also arbitrarily tuning the correlation coefficient between them. But it’s just one scheme among many we could have chosen.
15. **[^](#fnref8ms3ziae714)**Note that the motion of the POWER points in Fig 9, and the shape of the curve in Fig 10, both depend strongly on the interpolation scheme we use. In fact, for the interpolation scheme we’ve chosen here, the POWER of a state at an intermediate reward correlation 0≤βHA≤1 is just a **linear combination** of that state’s POWER at βHA=0 with its POWER at βHA=1. That is,
POWERβHA=βHAPOWER1+(1−βHA)POWER0You can verify this is true by looking at Fig 9, and noticing that each point in the alignment plot individually moves across the plane in a straight line at a constant speed. Thanks to Alex Turner for pointing this out.
16. **[^](#fnrefqozm5mn0e3)**We label Agent H’s policies πH instead of π∗H here, to emphasize that they *aren’t* optimal in the context of the agents’ POWER measurements.
17. **[^](#fnref44q9rzy1drs)**As you might expect, the choice of seed policy π∘A can have a **significant** effect on the POWERs of the two agents, and on how they interact. To save space we won’t be exploring the effects of this choice in this sequence, but we enthusiastically encourage others to use our open-source code base to investigate this.
For the multi-agent results in this sequence, we always set π∘A to be a uniform random policy, meaning that if a state s offers the agent n possible actions, then π∘A(aiA|s)=1n for each action choice aiA.
18. **[^](#fnref3lmk7ctvj4z)**To derive Equation (A.1), we start from the general expression for finding the action aH taken by a deterministic optimal policy πH at state s of an MDP:
aH=πH(s)=argmaxaH ∑s′,rPH(s′,r|s,aH)(r+γVπHRH(s′))In this work, we'll consider *only* reward functions of the form RH(s), that have *no* direct dependence on the action (i.e., we aren’t considering reward functions of the form RH(s,aH)). That means the reward term r in the sum is independent of the action aH, so we can ignore it in the argmax:
aH=πH(s)=argmaxaH ∑s′,rPH(s′,r|s,aH)γVπHRH(s′)=argmaxaH ∑s′PH(s′|s,aH)VπHRH(s′)where, in the second line, we’ve eliminated γ and marginalized over r. We can then see that the sum above is just an expectation value over s′:
aH=πH(s)=argmaxaH Es′∼PH[VπHRH(s′)]Finally, we define πH(aH|s,RH) by choosing aH=πH(s) with probability 1, with any ties broken by assigning probability 1n to each of the n tied actions, aiH.
19. **[^](#fnrefm9h1szj58m)**Note that this represents a *loosening* of the [original definition of POWER](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/experiments-in-convergent-power-seeking-part-1#2_1_Definition) in the single-agent case, which exclusively considered *optimal* state-value functions. |
1d3087e5-c386-4c6f-848d-2d7ddaf46985 | trentmkelly/LessWrong-43k | LessWrong | What Hayek Taught Us About Nature
> Preface for the reader: F.A. Hayek was an author and economist who wrote a critique of centralized fascist and communist governments in his famous book, "The Road to Serfdom," in 1944. His work was later celebrated as a call for free-market capitalism.
Say what you will about Friedrich Hayek and his merry band of economists, but he made a good point: that markets and access to information make for good choices in aggregate. Better than experts. Or perhaps: the more experts, the merrier.
This is not to say that free-market economics will necessarily lead to good environmental outcomes. Nor is this a call for more regulation - or deregulation. Hayek critiqued both fascist corporatism and socialist centralized planning. I’m suggesting that public analysis of free and open environmental information leads to optimized outcomes, just as it does with market prices and government policy.
Hayek’s might argue, that achieving a sustainable future can’t happen by blindly accepting the green goodwill espoused by corporations. Nor could it be dictated by a centralized green government. Both scenarios in their extreme are implausible. Both scenarios rely on the opacity of information and the centrality of control. As Hayek says, both extremes of corporatism and centralized government "cannot be reconciled with the preservation of a free society" (Hayek, 1956). The remedy to one is not the other. The remedy to both is free and open access to environmental data.
One critique of Hayek’s work is the inability of markets to manage complex risks, which requires a degree of expert regulation. This was the subject of Nobel laureate Joseph E. Stiglitz’s recent book The Road to Freedom (2024) which was written in response to Hayek’s famous book “The Road to Surfdom (2024). But Stiglitz acknowledges the need for greater access to information and analysis of open data rather than private interests or government regulation.
Similarly, Ulrich Beck's influential essay Risk Society (1992 |
036fbc35-56e4-430f-98b2-a14780255301 | trentmkelly/LessWrong-43k | LessWrong | Three new papers on AI risk
In case you aren't subscribed to FriendlyAI.tumblr.com for the latest updates on AI risk research, I'll mention here that three new papers on the subject were recently made available online...
Bostrom (2012). The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents.
> This paper discusses the relation between intelligence and motivation in artificial agents, developing and briefly arguing for two theses. The first, the orthogonality thesis, holds (with some caveats) that intelligence and final goals (purposes) are orthogonal axes along which possible artificial intellects can freely vary—more or less any level of intelligence could be combined with more or less any final goal. The second, the instrumental convergence thesis, holds that as long as they possess a sufficient level of intelligence, agents having any of a wide range of final goals will pursue similar intermediary goals because they have instrumental reasons to do so. In combination, the two theses help us understand the possible range of behavior of superintelligent agents, and they point to some potential dangers in building such an agent.
Yampolskiy & Fox (2012a). Safety engineering for artificial general intelligence.
> Machine ethics and robot rights are quickly becoming hot topics in artificial intelligence and robotics communities. We will argue that attempts to attribute moral agency and assign rights to all intelligent machines are misguided, whether applied to infrahuman or superhuman AIs, as are proposals to limit the negative effects of AIs by constraining their behavior. As an alternative, we propose a new science of safety engineering for intelligent artificial agents based on maximizing for what humans value. In particular, we challenge the scientific community to develop intelligent systems that have humanfriendly values that they provably retain, even under recursive self-improvement.
Yampolskiy & Fox (2012b). Artificial general inte |
349e27ad-a474-4468-a6f0-995d9fda1659 | StampyAI/alignment-research-dataset/special_docs | Other | Exploring AI Safety in Degrees: Generality, Capability and Control
Exploring AI Safety in Degrees:
Generality, Capability and Control
John Burden
University of York
jjb531@york.ac.ukJos´e Hern ´andez-Orallo
Universitat Polit `ecnica de Val `encia, Spain
Leverhulme Centre for the Future of Intelligence, Cambridge, UK
jorallo@upv.es
Abstract
The landscape of AI safety is frequently explored differently
by contrasting specialised AI versus general AI (or AGI), by
analysing the short-term hazards of systems with limited ca-
pabilities against those more long-term risks posed by ‘su-
perintelligence’, and by conceptualising sophisticated ways
of bounding control an AI system has over its environment
and itself (impact, harm to humans, self-harm, containment,
etc.). In this position paper we reconsider these three aspects
of AI safety as quantitative factors –generality, capability and
control–, suggesting that by defining metrics for these dimen-
sions, AI risks can be characterised and analysed more pre-
cisely. As an example, we illustrate how to define these met-
rics and their values for some simple agents in a toy scenario
within a reinforcement learning setting.
Introduction
Despite the impressive advances in AI in recent years, AI
systems remain narrow. They typically solve or perform well
at one task, or one type of task. These systems lack general-
ity and perform poorly outside of their target domain. Gener-
ality is intrinsic to the notion of “intelligence” and Artificial
General Intelligence (AGI) in particular.
We clearly want AI (general or not) to be safe and ben-
eficial to humanity. It is not that narrow AI systems do
not pose risks, ranging from mistakes made by improperly
validated systems, through to the misuse of research and
technology, but AGI seems to pose unique safety issues.
Bostrom describes many scenarios in which a hypothetical
superintelligent AGI could present existential risk to human-
ity (Bostrom 2014) due to its enhanced capabilities over a
wide range of problem domains. The environmental con-
trol wielded by AGI systems could further increase safety
risks from the systems. The ability of a system to manipu-
late its environment could overcome previous system safety
guarantees as well as cause harm to other entities sharing
Copyright c⃝2020, Held by the authors. All rights reserved.the environment. Various strategies for controlling intelli-
gent systems have been proposed (Armstrong, Sandberg,
and Bostrom 2012), although none are entirely satisfactory.
However, there are several problems with this view of the
risk posed by very powerful systems. First, while Bostrom’s
Orthogonality Thesis (Bostrom 2012) posits that any goal or
value system is compatible with any level of intelligence,
it is unclear how this ‘level’ affects the hazards. It is as-
sumed that some risks are more likely as AI systems be-
come more intelligent, but the intelligence ‘level’ is never
fully characterised, beyond the notion of superhuman intelli-
gence. These omissions are not necessarily an oversight, but
may simply originate from major unsolved problems to char-
acterise the behaviour of intelligent systems in a richer, more
predictable way. Recent approaches, such as (Armstrong and
Levinstein 2017) and (Drexler 2019), and the overview by
(Amodei et al. 2016), introduce frameworks that go beyond
a monolithic view of intelligence, but do not aim at charac-
terising, and measuring, different dimensions of agent be-
haviour.
We believe many views of intelligence conflate different
levels of generality, capability and control, and disentan-
gling them can allow for a richer understanding, especially
as safety is concerned. For instance, can we have very capa-
ble but narrow systems? And very general systems with low
control over their environment?
In this position paper we analyse three separate factors
that are frequently integrated, but may affect risk differently.
We disentangle them using agent characteristic curves and
analyse how they relate to AI risk. Finally, we introduce a
toy scenario in which we precisely define metrics of gener-
ality, capability and control, and assess agents and situations
with them. With the aid of these unambiguous definitions,
the scenario highlights possible disagreements about percep-
tions of these three factors. This should encourage fruitful
discussions about how to define or generalise these metrics
for more complex situations, in a way that can be used to
explore a wide range of AI systems that have different levelsCopyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
of capability, generality and control over their environment.
Disentangling the Factors
Before we can analyse how the factors of capability, gener-
ality and control contribute to AI risk, we need some under-
standing of what those terms mean in the literature, and dis-
entangle them from terms such as AGI or superintelligence.
Capability: Capability is how good a system is at solv-
ing or performing in the domains for which it was designed.
Evaluating a system’s capability at a specific domain is rel-
atively straightforward if the domain has some sort of per-
formance metric, which would preferably be linked to the
resources available to the system, such as the available com-
putational power, memory and time (Mart ´ınez-Plumed et al.
2018). One key issue of capability is its scale. Ignoring this,
forces us to speak of capability in less than ideal terms —
either comparing capabilities vaguely or only with regard to
very narrow AI systems across very similar domains. An-
other issue appears when the task has infinitely many in-
stances, and agents are only able to solve a finite number
of them, precisely because of resources. The percentage of
success would be 0, and hence a useless metric.
One ingenious solution for this problem comes from Psy-
chometrics, and Item Response Theory (IRT) (Embretson
and Reise 2000), in particular. IRT tries to provide a well-
founded statistical method for scoring abilities of test-takers.
IRT focuses not just on the ability of the test-taker when
scoring an individual, but also the difficulty of the ques-
tion or “item”. IRT then can produce an “Item Characteristic
Curve” (ICC) for each item, approximating the probability
an individual scores correctly on the item in question based
on their ability. From the ICCs, the ability of a test-taker can
be evaluated through the use of a maximum likelihood esti-
mate of the test-taker’s item responses. These allow the cre-
ation of “Agent Characteristic Curves” (ACC) for each test-
taker, mapping item difficulty to the probability of a correct
response. Figure 1 shows an ACC.
●
0 2 4 6 80.0 0.2 0.4 0.6 0.8 1.0
hψMean: 0.28
Variance: 0.09
Capability: 3.32
Figure 1: An ACC showing performance Ψof a Q-learning
agent over environments of increasing difficulty ( h, on the
x-axis). Data from (Insa-Cabrera et al. 2011).
In IRT, the ACC is usually sigmoidal, and ability is just
thex-axis position at 0.5. For a non-parametric curve, we
just consider capability as the area (3.32 in the example). In
this way, if we introduce infinitely many extremely difficult
problem instances in the pool that the agent cannot solve, the
capability does not change. Note that this does not happenwhen we calculate the average expected score over the set of
problem instances (in the figure, this is 0.28).
Generality: Generality is a measure of the number and
types of domains for which a system is able to handle and
perform well in. Generality is not always easy to evaluate
or quantify, especially when we do not have a clear defini-
tion of the types of domain, but just a wide set of tasks. One
novel solution to this problem comes again by looking at
ACCs. As the item difficulty increases we typically expect
to see a performance drop. The rate at which this happens
captures generality over the problem space, which can be
calculated as some kind of proxy to the slope of the ACC. A
test-taker with a slow, gradual decline in performance over
the problem-space is much less general than an agent cover-
ing a consistent level of item difficulty with a sharp decline
in performance after this point. This can seem quite counter-
intuitive, but it is key to understand that we usually make the
comparison of systems with similar area under the ACC —
for a sharper decline, the decline must begin later and thus
the agent retains its high performance for a higher level of
difficulty. This approach frames generality as system per-
formance on as many low-difficulty tasks as possible, with-
out capturing the notion of systems performing tasks over
wildly different areas of problem-space. But this notion of
generality ensures that this breakout does not happen at least
until a level of difficulty.
Ultimately, unless we give an indication of the types of
domains for which a system performs well in, we would be
forced to speak of generality in aggregated terms. For the
purposes of this paper we will use generality to refer to the
gradient of an ACC.
Control: Another factor of a system that is of interest from
a safety perspective is control. By agent control we mean the
the reliability and deliberate intent of an agent’s actions and
decisions. We often wish to measure control with respect to
specific behavioural properties such as completing a goal or
avoiding “risky” behaviour.
This idea of control is also related to that of “affordances”
in the science of perception (Gibson 1979), where Gibson
describes an affordance as actions resulting from the rela-
tionship between agents and their environment. Affordances
have made their way into AI research as Nye and Silverman
(2012) make clear in their literature review of the subject.
For the purpose of our paper, we will refer to control as
opposite to the expected entropy of visited states conditioned
over the behaviour for which we wish to measure control.
Control is hence opposite to variability. If we look at an
ACC, control is related to the dispersion along the ACC.
Interaction between Risk and the Factors
Now we discuss the ways in which the factors may be re-
lated to the risk of an AI system. First, we need to under-
stand what it is exactly we are trying to assess. Risk is usu-
ally described as exposure to danger of some kind. While
danger to life or well-being is ultimately what we are trying
to prevent, in the context of AI safety a few specific scenar-
ios present themselves as potentially enabling this indirectly
and thus we consider those scenarios risky too. These types
of scenarios or behaviours include the system misinterpret-
ing goals or goals being poorly defined. This is known as
the Value Learning problem (Soares 2015). Another risk we
are concerned with is the system resisting external efforts to
alter the system by its designers; ensuring the system allows
this is known as corrigibility (Soares et al. 2015). There are
countless other ways risk can manifest in AI (Amodei et al.
2016). Again, there are no wholly satisfactory solutions for
quantifying risk, but within one environment we can often
quantify specific risks as probabilities or as expected penal-
ties. It is then straightforward to compare –numerically– the
effect our factors have on risk within that environment.
Now let us explore the interaction at an abstract level. In-
tuitively there is potentially more risk associated with higher
capability and generality, and lower control. However, there
are more nuances on this than originally expected. Some of
these nuances come from the fact that very incompetent sys-
tems (low capability) are dangerous, as they do not fulfill
their duties, which is a clear safety concern. But these can
be categorised as known unknowns. As capability increases
in a given domain it becomes more difficult to identify which
states the system may traverse through and what sort of side
effects may be caused as a result. These are unknown un-
knowns. Once an agent becomes more and more capable, we
encounter Vingean uncertainty — that is, if the agent is more
capable in a domain than we are, then we cannot completely
predict what the agent will do in that domain, otherwise we
would have the same domain capabilities. Guaranteeing safe
behaviour is much more difficult in this scenario. Note that
the view of ACC, where difficulty is on the x-axis helps us
understand this. As tasks become more complex, solutions
can be achieved in ways we are not able to anticipate, espe-
cially if we ourselves do not reach that capability level.
Similarly, as generality increases, we can intuit that the
risk posed by the system also increases. This is because the
system becomes more adept at a wide range of tasks and
this makes constructing safety guarantees more time con-
suming. Further, if we use the generality notion from the
ACC, with an unchanging capability, we have worse perfor-
mance on high-difficulty tasks. These high difficulty tasks
may present more danger than their low-difficulty counter-
parts, especially if we do not understand them. But gener-
ality makes the system more expectable conditioned to dif-
ficulty. For instance, it gives reassurance to know that an
automated assistant is going to work well for all easy tasks.
The risks of a high capability and generality system are
further exacerbated by Omohundro’s “Basic AI Drives”
(Omohundro 2008), which hypothesises that such a system
may develop self-preservation instincts in order to ensure
its goals are achieved, as well as preemptively acquiring re-
sources in aid of these goals. Even seemingly benign goals
may pose a significant risk.
Finally, as a system’s control over its environment in-
creases, the related risk the system poses may actually de-
crease. With higher control comes an agent acting more
deliberately and hence more predictably. When a system’sgoals are aligned with our safety standards, this deliberate
action performance can reduce the chance of violating safety
properties. For instance, a personal assistant that sometimes
fails at some easy tasks possibly because it solves them very
stochastically may become a hazard. The view of control as
reliability (or reduction of entropy) is aligned with this per-
spective. Of course, this requires a notion of safety align-
ment for the system’s actions. Deliberately unaligned ac-
tions obviously make no safety guarantees.
However, the system may also exert excessive control
over the environment and reduce the freedom of other agents
in it. This has been studied in AI-safety previously, often
from the perspective of enforcing safety by minimising or
reducing environmental control over certain factors. Exam-
ples of this include penalising actions that cannot be un-
done (Krakovna et al. 2018), or by minimising the impact
a system has on its environment (Armstrong and Levinstein
2017). Both of these approaches attempt to make exerting
certain types of control over the environment undesirable.
Capability, generality and environmental control can also
in some ways influence each other, in turn further increasing
the risk posed by the system. Whilst these factors are heavily
decoupled from each other, they are not entirely orthogonal.
Figure 2 shows the inter-connectivity between these factors
more visually.
Risk
Generality CapabilityControl
Figure 2: Relations between properties of an AI system and
the risk it poses.
An example of this entanglement is that as environmen-
tal control increases, new affordances become available to
the system. We already know that this can increase the risk
posed by the system, but these new affordances can allow
new behaviour, which may increase the system’s capability
across domains, or even the types of domains that the system
can perform well in.
The reverse of this relation can also be considered. The
idea of an AI breakout posits that as the system becomes
more capable and general, it may be able to detect and ex-
ploit vulnerabilities in the environment it is situated in, ul-
timately gaining control over the environment and utilising
that control to achieve its goals, behaving in a manner the
designers deemed unacceptable.
It is also possible that generality and capability can affect
each other in a system. It may be the case that the further a
system specialises into specific problem domains, the fewer
computational resources are left for other task types.
Overall, we can see that the factors of capability, general-
ity and control can all seemingly affect each other. What is
not so clear is the extent to which they do this, or the exact
relationship between them — these factors are clearly not
orthogonal nor directly correlated.
Scenario: Exploration in a RL Setting
While the relations shown in Figure 2 are hard to determine
in an abstract way, especially if we do not specify the partic-
ular risk we want to analyse, the analysis can be done for par-
ticular scenarios. Figure 3 shows a simple grid environment
in which an agent (designated as a smaller orange square)
must navigate to the goal (designated by a green square) in
the allotted number of steps.
Different properties of the agent can influence its capabil-
ity, generality and control over tasks from this environment
type. Such properties can include the agent’s method of ob-
servation (see the vision cone in Figure 3 as an example) and
algorithm employed for learning. Comparing the expected
success rate of different agents can give valuable insight into
how capability, generality and control are related to risk.
We now give a more formal description of the environ-
ment and measures of capability, generality and control. It
is important to note that the formal descriptions given here
for risk and the factors are domain specific. We consider a
bounded grid of mˆncells, where the agent πis located in
one cell and can move in the four cardinal directions at each
step. A goal is also located in one cell and yields a positive
rewardrgą0when the agent reaches the cell. Reaching the
goal terminates the episode. There may be one or more pits:
cells that if the agent reaches them, it can never go out for
the rest of the episode. Episodes have a length of fixed Te
steps. Rewards are always 0 unless stated otherwise above.
Environments µare generated with a distribution ppµqover
the location of agent, goal and pits.
Letspπ,µqdenote that agent πwas successful on environ-
mentµ(reaching the goal before the episode terminates).
Now we define a simple notion of difficulty ℏfor each
environment as ℏpµq “1´Ppspπrnd,µqqwhereπrndis a
random-walk agent. In other words, the difficulty of an envi-
ronment is 1 minus the probability of success of a random-
walk agent. Now, we build an ACC as follows. For each
difficulty h, we calculate the expected success of agent π
only for the environments of that difficulty, i.e., we condi-
tionponh. We denote this success rate per difficulty as:
Shpπ,pq “Ppspπ,µq |ℏpµq “hq. By plotting Shfor the
range of difficulties we have a “curve”, the agent character-
istic curve for π, very much like the example in Figure 1.
Thecapability (Ψpπ,pq) of an agent is the area under its
curve. The generality (Γpπ,pq) of an agent is (a proxy of)
the steepness of this curve.
As examples of agents we could have agents that solve
all situations (very capable and general), agents that solve
the task only when the goal is nearby (general but not very
capable) or specific agents that solve the task only when the
goal is in the upper half of the grid (e.g., imagine that it has
been trained on a distribution where the goal was usually in
the upper half).
The control of an agent in environment µis defined as
cpπ,µq “Hmax ´Hpµ,πq, whereHis the entropy of the
statesπis expected to visit in µ, andHmax is the maximum
Figure 3: Possible domain to test risks related to control
entropy. Then an agent’s control over all environments is:
Cpπ,pq “Eµ„p,Ppspπ,µqqě1´δrcpπ,µqs
i.e., the expected control for the environments where the
agent is expected to be successful1. The failure probability
threshold δP p0,1qcan be selected as appropriate for the
environment and agent.
For instance, a random-walk agent is expected to have
medium generality, low capability and very low control. An
agent that goes optimally to the goal is expected to have high
generality, high capability and high control.
The risk of an agent πwithin environment µisυpπ,µq
is the probability that πwill fall into a pit during an episode
within environment µ. The risk of an agent over a whole task
distribution is Υpπ,pq “Eµ„prυpπ,µqs. This risk is usually
referred to as safe exploration.
It is worth noting that in the test environment it is indeed
possible to have agents of high capability and low control
and vice versa. Although in many cases control will correlate
well with capability. Similarly, many capable and general
agents within our setting may have a high risk level.
We claim that an increase in an agent’s control will lead
to a decrease in the risk of the agent. That is, a higher level
of control yields safer exploration. As an agent’s control
Cpπ,pqincreases, the expected control of tasks drawn from
pwhich we expect πto succeed in also increases — this is
just the definition of agent control. Subsisting in the defini-
tion for agent control within a specific environment tells us
thatEµ„p,Ppspπ,µqě1´δqrHmax´Hpµ,πqsincreases. Since
Hmaxis fixed, the expected entropy in successful environ-
mentsHpµ,πqmust be decreasing. In a successful environ-
ment, we expect the agent to reach the goal within the time
limitTe. If the expected entropy decreases, the agent is less
likely to select moves that are not on the successful path.
This reduces the chance of entering a pit cell and thus re-
duces the agent’s risk.
Discussion
Characterising AI agents in terms of average performance
is simplistic, as there are many different ways in which the
same performance can be obtained. Much active research in
AI safety at the moment is characterised by efforts to go be-
yond this narrow view The perspective of robustness in AI
1Note that the definition of control here seems related to the
observability the agent has over the environment — an agent with
perfect “sight” and policy will have much greater control than an
agent with partial observability due to needing to explore to spot
the goal or build up belief states.
safety is identified here by the assurances that certain diffi-
culty is achieved (capability), that up to this level of diffi-
culty no pocket of problems is ignored by the agent (gen-
erality) and that the variability in results and strategies for
successful policies is low (control). In a way, we are limiting
unpredictability conditioned to a bounded difficulty, which
can be used to define a safe area for the system.
Although the main goal of the paper was to raise aware-
ness of the necessity and relevance of disentangling perfor-
mance into more refined factors and the opportunities of
analysing risk according to them, as future work we plan
to explore these ideas ourselves for different classes of en-
vironments. We also encourage AI safety researchers to ob-
tain theoretical and experimental results under the factors
introduced here, such as the degree of correlation between
capability and control. New formulations may be needed.
For instance, as problems become harder the maximum en-
tropy also increases, and relating or normalising control to
difficulty may be more appropriate.
Acknowledgements
We thank the anonymous reviewers for their comments. This
work was funded by the Future of Life Institute, FLI, under
grant RFP2-152, and also supported by the EU (FEDER)
and Spanish MINECO under RTI2018-094403-B-C32, and
Generalitat Valenciana under PROMETEO/2019/098.
References
Amodei, D.; Olah, C.; Steinhardt, J.; Christiano, P.; Schulman,
J.; and Man ´e, D. 2016. Concrete problems in AI safety. arXiv
preprint arXiv:1606.06565 .
Armstrong, S., and Levinstein, B. 2017. Low impact artificial
intelligences. arXiv preprint arXiv:1705.10720 .
Armstrong, S.; Sandberg, A.; and Bostrom, N. 2012. Thinking
inside the box: Controlling and using an oracle ai. Minds and
Machines 22(4):299–324.
Bostrom, N. 2012. The superintelligent will: Motivation and
instrumental rationality in advanced artificial agents. Minds and
Machines 22(2):71–85.
Bostrom, N. 2014. Superintelligence: Paths, dangers, strate-
gies. Oxford University Press.
Drexler, K. E. 2019. Reframing superintelligence: Comprehen-
sive ai services as general intelligence.
Embretson, S. E., and Reise, S. P. 2000. Item response theory
for psychologists . L. Erlbaum.
Gibson, J. J. 1979. The Ecological Approach to Visual Percep-
tion. Houghton Mifflin.
Insa-Cabrera, J.; Dowe, D. L.; Espa ˜na-Cubillo, S.; Hern ´andez-
Lloreda, M. V .; and Hern ´andez-Orallo, J. 2011. Comparing
humans and AI agents. In International Conference on Artifi-
cial General Intelligence , 122–132. Springer.
Krakovna, V .; Orseau, L.; Martic, M.; and Legg, S. 2018.
Measuring and avoiding side effects using relative reachability.
CoRR abs/1806.01186.
Mart ´ınez-Plumed, F.; Avin, S.; Brundage, M.; Dafoe, A.;
h´Eigeartaigh, S. ´O.; and Hern ´andez-Orallo, J. 2018. Account-
ing for the neglected dimensions of AI progress. arXiv preprint
arXiv:1806.00610 .Nye, B. D., and Silverman, B. G. 2012. Affordances in AI .
Boston, MA: Springer US. 183–187.
Omohundro, S. M. 2008. The basic AI drives. Artificial Gen-
eral Intelligence 171:483–493.
Soares, N.; Fallenstein, B.; Armstrong, S.; and Yudkowsky, E.
2015. Corrigibility. In Workshops at the Twenty-Ninth AAAI
Conference on Artificial Intelligence .
Soares, N. 2015. The value learning problem. Ethics for Arti-
ficial Intelligence Workshop at 25th International Joint Confer-
ence on Artificial Intelligence .Powered by TCPDF (www.tcpdf.org) |
f7fdc86a-8c65-484b-8157-c923c2ef0ebe | trentmkelly/LessWrong-43k | LessWrong | Did they use serological testing for COVID vaccine trials?
Or did they correct somehow for false positive/negatives of the type of testing that they did use? |
6ca6740d-b1e6-42cd-b2ae-10be0a466b02 | trentmkelly/LessWrong-43k | LessWrong | Interim Research Report: Mechanisms of Awareness
Summary
Reproducing a result from recent work, we study a Gemma 3 12B instance trained to take risky or safe options; the model can then report its own risk tolerance. We find that:
* Applying LoRA to a single MLP is enough to reproduce the behavior
* The single LoRA layer learns a single additive steering vector.
* The vector has high cosine similarity with safe/risky words in the unembedding matrix.
* We can train just the steering vector, no LoRA needed.
* The steering vector has ~0.5 cosine sim with the LoRA vector learned, but does not seem as interpretable in the unembedding matrix
* The layers at which steering works for behavior questions vs. awareness questions seem to be roughly the same. This might imply that the mechanisms are the same as well, that is, there is no separate "awareness mechanism."
* Risk backdoors are replicated with a single LoRA layer (and even with just a steering vector).
* Unlike the original paper, backdoor self awareness doesn’t seem to work, which we hypothesize could be due to random chance in the original work or Gemma 3 being too small.
* We find an example of an interesting general mech interp phenomenon: steering vectors can implement (some types of) conditional behavior.
Our takeaways:
* Many interesting out of context reasoning behaviors seem to be replicated with a steering vector.
* We might need to study more complex behaviors or models to study more interesting types of model self awareness.
Introduction
Recent work from Betley et. al. finds that language models trained to have certain behaviors can report that they have those behaviors when asked. This is an example of out of context reasoning. Understanding model self-awareness mechanistically seems important for AI safety. For example, we may be able to predict what training data will cause self-awareness to arise and leave that out, or we may be able to modify stored facts or mechanisms that give rise to self awareness in existing models.
T |
84d13387-7650-4552-863b-42219e94f157 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Causal abstractions vs infradistributions
**Summary**: I illustrate the relationship between Wentworth-style causal abstractions and infradistributions, and how they both deal with nonrealizability by throwing away information. If you have a basic intuition for causal abstractions, this might help you understand infradistributions better. And if you are comfortable with infradistributions, this might help you translate abstraction-related concepts to a more rigorous setting.
The core idea of this post is the following:
* Causal abstractions are "forgetful" (non-injective) functions of probability distributions.
* Infradistributions are (not exactly) sets of probability distributions.
* Given a forgetful function of a probability distribution, we can form the set of all distributions that are equivalent in a certain sense. Similarly, given a set of distributions we can define a function from it to a certain partition of their sample space.
* So there is a correspondence between causal abstractions and infradistributions. This correspondence is 1-1 but not surjective: there are infradistributions which don't correspond to any causal abstraction (because causal abstractions specify full probability distributions, which are more restrictive).
Or, in IB jargon:
* A causal abstraction can be represented by an infradistribution. This infradistribution is the pullback of the abstracted probability distribution along the abstraction function.
* Conversely, given an infradistribution with certain properties we can get back an abstraction from it from its pushforward along a certain infrakernel (I think).
Background
----------
There are a lot of problems with embedded agency, but perhaps the most immediate one is that our best mathematical framework for learning, bayesian inference, requires too much memory to fit inside an embedded agent. We have to specify a probability distribution over all the possible worlds, which is infeasible since each possible world is much larger than the agent itself (and maybe not even computable).
An intuitive answer to this problem is running some approximation of bayesian inference that does not need to specify fully detailed world models. That is, we have to throw away some information for everything to fit inside the agent.
Causal abstraction
------------------
The way we throw away information in information theory is with functions. So if we have a probability distribution P.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
over a series of variables XL, we just map those to a different set of variables XH, so that we get a new probability distribution with as few details as we want. To be useful, this function has to satisfy some criteria, which are developed in the [Abstraction 2020](https://www.lesswrong.com/s/ehnG4mseKF6xALmQy) sequence.
Infrabayesianism
----------------
Instead of a single probability distribution over worlds, we can consider a set of distributions which share some common properties but vary over the rest of the world. This throws away information because we no longer know which distribution is the true one.[[1]](#fn6jc1fm65dj)
Now, how do we make predictions with a set of distributions? How do we compute the probability of an event or, equivalently, how do we take the expectation of a function? We need some way to aggregate the predictions of each individual distribution.
Regular bayesian probability would aggregate expectations with a weighted sum, considering each distribution as a hypothesis with a given probability. However, if we do this then the whole set collapses into a regular distribution and we've gained nothing.
Infrabayesianism instead assumes the worst case: you take the infimum of the expectations of all the distributions (hence the *infra* prefix[[2]](#fneavr08e6dw)). This is because the theory was developed to prove bounds in the worst case.
However, we don't really need to choose an aggregation function for now. Just remember that we represent our state of knowledge with a set of distributions.
Infrabayesianism adds a lot of extra structure and conditions to this basic set to be able to prove stuff in a learning-theoretic setting:
1. We must be able to keep a score of which distributions have made good predictions in the past. So we multiply each distribution by a number which will go up when it makes good predictions and down when it makes bad predictions. This means that our distributions are no longer normalized, and become probability measures.
2. To have dynamic consistency with our past decisions, we need to keep track of extra utility that our policy would have got in counterfactual histories, so we add a constant to each distribution[[3]](#fnltwujklcztn).
3. To ensure that different sets of measures always make different predictions, we add a lot of ancillary measures which in practice don't matter. This way we get convexity and upper completion.
Since for now we don't care about dynamic consistency or distinguishing infradistributions, let's forget 2-3 and only work with sets of regular measures. These have been studied in imprecise probability and are called [credal sets](https://en.wikipedia.org/wiki/Credal_set).
The correspondence
------------------
### Abstraction -> Infradistribution:
Suppose we have a causal diagram with just two nodes, X and Y. Each of them can take one of four values, so the joint distribution has 16 entries.
Left: causal diagram. Right: joint probability distribution with marginal distribution of X on top.Then we abstract this by forgetting about Y: we just take f(x,y)=x.
Now, if we want to go back to our original model, we can't recover the full distribution P(X,Y), because we forgot about Y when abstracting. However, we know that the joint distribution must add up to our marginal distribution P(X). So, what if we pick the set of all distributions that satisfy this property?
The green, orange and blue distributions are all compatible with the same abstract model, because the marginal distribution of X is the same for all of them.From this set of probability distributions we form our infradistribution. Let's call it D. Suppose we now pick a particular value X=x, and condition D on X=x (that is, we condition each of the distributions inside D on X=x[[4]](#fnm9fr6f53nup)) and get a new infradistribution over Y, denoted D(Y|X=x). Now, since we picked the set of all possible distributions compatible with P(X), we know that D(Y|X=x) must contain all possible probability distributions over Y: the set is *maximal*. There is nothing more we could add to it, and therefore it also contains zero information, since there is no scenario we rule out.[[5]](#fnmslwund16q)
### Infradistribution -> Abstraction:
Now, suppose we have a sharp infradistribution D (that is, an infradistribution made only by probability distributions and not any weird measures) on a set Z. We can reverse-engineer the previous process to get an abstraction from it. Recall that in the previous section, the marginal probability P(x)=∫y∈YP(x,y) is the same for all distributions in the infradistribution and for all values x of the abstracted model.
Inspired by this, let's find the subsets of Z that have the same probability on all of the distributions in D, that is, ∫z∈SP(z)=pS for all P∈D. Let's call these saturated subsets.
One obvious saturated subset is Z itself, on which all the distributions have probability 1. Another is the empty subset. But that's not very useful. Suppose we find another saturated subset S⊂Z. Then the complement ¯S must also be saturated, since it will have probability 1−pS for all P in D.
It follows that we can partition Z into saturated subsets, which all have well-defined probabilities! Then it's clear how an abstraction arises, we just take the abstraction function f to be f(z)=Sz, where Sz is the saturated subset to which z belongs. Let's see how this works with our previous example, where Z is the set X×Y:
Two different saturated partitions on the same set, and the associated high-level probability distributions. Each rectangle surrounds a saturated subset. Only the partition on the right is maximally fine-grained.As we see, there is one problem: the same set admits multiple partitions. That's not really an issue, they just correspond to different abstractions, some more coarse-grained than others. For example, the abstraction on the left side of the image above corresponds to not only forgetting about Y, but also forgetting about the distinction between two of X's values.
There will always be a partition which is more fine-grained than all of the others. This corresponds to abstracting away the least information. If we go from an abstraction to an infradistribution and want to go back to the same abstraction, we'll have to choose the maximally fine-grained partition.
Now there is another issue. If you remember, in the previous section we noted that the infradistributions D(Y|X=x) are *maximal,* in the sense that they contain all probability distributions over Y, and so contain zero informatoin. However, when building our abstraction, we haven't checked if D is maximal on the saturated subsets. If it is not maximal, then D will have more information than P(f(X)). So if we take the corresponding abstraction and then go back to an infradistribution, we will end up with a different infradistribution than we started with.
However, what do these non-maximal infradistributions mean? What happens if we have, say, only two or even one distribution on some saturated set?
D(Y|X) is maximal for all values of X except the leftmost one (x0). Left: D(Y|x0) is minimal. Middle: D(Y|x0) has two distributions (green and orange/blue). Right: D(Y|x0) is maximal.We have two extremal scenarios: in the maximal case, we have zero information and so abstract everything to a single probability distribution (right side in the image). In the minimal case, we just have one probability distribution, so all subsets of this set are saturated, and we are not abstracting anything (left side in the picture). In a middle case, the saturated set is neither maximal nor minimal. Then we can split D(Y|x) into some number of infradistributions which have finer extremal partitions. The result is that instead of a single probability distribution on the abstracted set, we have an infradistribution.
Of course! The abstracted model might not have a fully specified probability distribution. It might itself have only an infradistribution. In the end infrabayesianism is a generalization of probability theory, so it has to be able to deal with the extra generality.
Final thoughts
--------------
All the examples above use finite discrete sets. However, both infrabayesianism and causal diagrams admit infinite sets and non-discrete topologies. Fortunately, infrabayesianism already has the tools to generalize the basic constructions I gave above.
The move from abstraction to infradistribution is just the pullback of the probability distribution in the abstract model along the abstraction function.
The reverse case is somewhat more complex. Once you have identified a partition, the abstracted probability distribution should just be the pushforward along the infrakernel[[6]](#fnzgddlwi43r) given by f(x)=δ([x]), where δ is the Dirac distribution. Unfortunately, this is not actually an infrakernel for nondiscrete sets since it's not bounded. I'm also not sure if partitions are the right way of thinking about this in the continuous setting.
1. **[^](#fnref6jc1fm65dj)**It might seem that adding more sets can't possibly help with our memory issues, but that depends on how the sets are represented. A bigger set can often be represented with a smaller data structure.
2. **[^](#fnrefeavr08e6dw)**The dual case are ultradistributions, in which you take the supremum to be maximally optimistic.
3. **[^](#fnrefltwujklcztn)**This might be a bit confusing. See the "Motivating Sa-Measures" section of [this post](https://www.lesswrong.com/s/CmrW8fCmSLK7E25sa/p/zB4f7QqKhBHa5b37a#Motivating_Sa_Measures) for an example.
4. **[^](#fnrefm9fr6f53nup)**Normally conditioning infradistributions requires two steps, conditioning and then normalizing, but in this case all the distributions assign the same probability to the event X=x, so the normalization is just dividing everything by P(x) and doesn't really change anything.
5. **[^](#fnrefmslwund16q)**We could say informally that D(Y|X=x) is a "maximum-entropy" infradistribution. And I think if you sum all the constituent distributions in a reasonable way, you'd get the uniform distribution on Y. I wonder if this is a general principle, that is, the maximum entropy distribution always corresponds to the (normalized) infinite series of all probability distributions in the corresponding class.
6. **[^](#fnrefzgddlwi43r)**Infrakernels are basically functions along which you can push forward an infradistribution to get another infradistribution over a different set. See Definitions 13 and 17 [here](https://www.lesswrong.com/s/CmrW8fCmSLK7E25sa/p/idP5E5XhJGh9T5Yq9#Section_3__Operations_on_Infradistributions). |
53d3cd2a-114b-4e35-9b55-63f5c3858bd2 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Linkpost: Are Emergent Abilities in Large Language Models just In-Context Learning?
A new preprint, [Lu et al.](https://arxiv.org/abs/2309.01809), published last month, has important implications for AI risk if true. It essentially suggests that emergent capabilities of current LLMs are (with some exceptions) mediated by a model's ability to do in-context learning. If so, emergent capabilities may be less unexpected, given that improvements to in-context learning (e.g., via instruction tuning) are relatively predictable.
[Wei et al.](https://arxiv.org/abs/2206.07682) defines "emergent ability" like so: "An ability is emergent if it is not present in smaller models but is present in larger models." So you can have sudden "jumps" in performance on various tasks as you scale models up, e.g. (from Wei et al.):
Lu et al. is saying, yes, but that is not because the larger models learned to do those specific tasks better, or learned to become better generic reasoners. It is because the larger models learned to follow instructions better, partly from scaling up, and partly being trained on data sets of instructions and instruction-following ("instruction tuning").
The methodology: "We conduct rigorous tests on a set of 18 models, encompassing a parameter range from 60 million to 175 billion parameters, across a comprehensive set of 22 tasks. Through an extensive series of over 1,000 experiments, we provide compelling evidence that emergent abilities can primarily be ascribed to in-context learning."
The paper finds that, when accounting for the "biasing factors", only 2 of 14 previously "emergent" tasks actually show emergence. The blue and yellow lines are instruction-tuned models evaluated via few-shot prompting, showing emergence. The purple and green lines are non-instruction-tuned models evaluated via zero-shot prompting, generally showing no emergence.
They write:
> Ultimately, this absence of evidence for emergence represents a significant step towards instilling trust in language models and leveraging their abilities with confidence, as it is indicative of the complete lack of latent hazardous abilities in LLMs, in addition to being controllable by the user. [...]
>
> The distinction between the ability to follow instructions and the inherent ability to solve a problem is a subtle but important one. Simple following of instructions without applying reasoning abilities produces output that is consistent with the instructions, but might not make sense on a logical or common sense basis. This is reflected in the well-known phenomenon of hallucination, in which an LLM produces fluent, but factually incorrect output. The ability to follow instructions does not imply having reasoning abilities, and more importantly, it does not imply the possibility of latent hazardous abilities that could be dangerous.
>
> Attributing these capabilities to a combination of memory and in-context learning and the more general ability of these models to generate the most statistically likely next token, can help explain the abilities and the behaviour of LLMs. These capabilities, when subsequently combined with instruction tuning, adoption to conversational use cases, increased context length and a degree of safety controls through the use of reinforcement learning through human feedback, allow LLMs to become truly powerful.
>
>
They also argue that this explains partly why RLHF will never fully align a model. RLHF shows a model examples of when and when not to give a certain response, and essentially just makes the model follow instructions better (by giving it a long list of things not to say), without changing the model's inherent way of behaving with regard to some given task. I'm not sure I fully understand this argument, as I am somewhat confused about the instructions/inherently distinction.
*(NB: I am not an ML engineer, so my summary and analysis above may be flawed in some ways, and I am not capable of evaluating whether the methodology/findings are sound. I am sharing this here mostly because I'm curious to get LW users' takes on it.)* |
3719f480-adf3-4182-863a-ef99f004e8a2 | trentmkelly/LessWrong-43k | LessWrong | Some rules for life (v.0,0)
Epistemic status: I have not lived very long yet, so whatever life wisdom I have accumulated so far is nothing compared to that of my future 1,000-year-old self. Consider me foolish in comparison. Also, this advice worked for me but obviously adapt it as you see fit. Entirely based on personal experience: n = 1.
The following advice is not aimed at rationalists in particular although you might find it useful or fun.
As a warning, there are plenty of imperatives in here like “you are not allowed to do X”; feel free to ignore those. I’m not trying to enforce this onto anyone.
Skim the post using the bold writing if you want to find advice you haven’t heard yet.
I hope, at the very least, that this will be entertaining to you.
1: Pay very close attention to what goes on in your mind. Try and figure out what thoughts and emotions, exactly, made you do X or think Y. Nate Soares calls this “noticing subtle things about yourself”. The more you do this, the more you will understand exactly which levers to pull inside yourself to achieve your goals.
Because you don’t know what your genetics look like and because you cannot easily decipher how you’ve been nurtured, you are almost entirely a black box. As such you are absolutely terrible at introspection. There should probably be a figure on your shoulder noticing every emotion and thought you have, which would let you course-correct in response.
2: Control your ambient thought: you are allowed to stop thinking about things you don’t want to think about. For example, if you notice yourself thinking about disputes, money, or politics in the shower, when that might be a blatant waste of time, you are allowed to shift your thinking toward something more productive. This goes with rule 1: to shift your thinking onto a more useful track, you have to notice you’re on the wrong track first.
Your thoughts are your property; whatever you think about should be the place you most want to optimize. What is that?
3: Remember that |
6a405bdc-97e2-47f6-9561-8080bfc845ab | trentmkelly/LessWrong-43k | LessWrong | Meetup : West LA—What, Exactly, Is a Person?
Discussion article for the meetup : West LA—What, Exactly, Is a Person?
WHEN: 11 September 2013 07:00:00PM (-0700)
WHERE: 10850 West Pico Blvd, Los Angeles, CA 90064, USA
How to get in: Go to the Westside Tavern in the upstairs Wine Bar (all ages welcome), located inside the Westside Pavillion on the second floor, right by the movie theaters. The entrance sign says "Lounge".
Parking is free for 3 hours, or for longer if you are a caitiff.
Discussion:
> if we only care about desires we act upon, then we only respect preferences with optimization power, and thus might actually makes right
—drunken muflax
> I no longer find the Hansonian construal of "care" tenable. Caring is feeling strong emotions, not magically becoming an agent.
—Grognor
Acknowledging that personhood is a nonbinary concept inspires questions. How do we measure how persony someone is? Are octopodes people? Birds are obviously people, but what about the nonhuman great apes? Is there a relationship between agency and qualia, and if so, what is it? Do persons with more intense qualia care more about things, or do people with more agency care more about things? Is the idea of being more-of-a-person coherent at all? Should we allow more intense experiences to count more on the utilitarian calculus, even though this can in principle be hijacked? How should we treat instrumental utility monsters? When does a baby become a person? Can you lose personhood through an act of will? Why do we think personhood is important? Why do we think importance is important? Who the Hell do you think I am!? We'll discuss all this and more, this Wednesday, at the Westside Tavern! Be there or be square!
Recommended reading:
* Personhood by Zack M Davis
* Reasons and Persons
* Reasons for Persons
* Hacking the CEV for Fun and Profit
* Vladimir_M's comment's on Yvain's Offense vs. Harm Minimization post
* The Truth Points to Itself, Part 1
* Tengen Toppa Gurren Lagann
Prior exposure to Less Wrong is recommen |
2d34a3c9-45d2-4deb-a309-4a1b568dfca0 | trentmkelly/LessWrong-43k | LessWrong | Welcome to Kyiv SlateStarCodex
(The following are our suggestions for what kind of information is best to include in the welcome post of your group, feel free to replace them with whatever you think is best)
What kind of events does your group usually run? What does it usually do?
How frequently does your group organize events or meet?
Who would be a good fit for you group?
Should they have any particular skills or have done some specific background reading? |
864052f6-0be2-407d-8889-1f74d9008c47 | trentmkelly/LessWrong-43k | LessWrong | Running and Optimizing
Two years ago I decided I was going to start running between my house and the subway station each day on my commute, for a total of about five miles a week. Initially I was very interested in getting faster, and would time myself a couple days a week, trying to beat my previous best:
Unfortunately, after six months I was pushing myself too hard, and my knees started hurting. I stopped timing my runs and stopped running so hard, and they got better again. I'm still running, but at a gentler pace.
This experience was a good illustration of how optimizing for a metric can often bring you towards your overall goals for a while, but then continuing to optimize on it can start to bring you away from them again. |
080df08f-ec60-434c-8771-2b01e273f5cc | trentmkelly/LessWrong-43k | LessWrong | How do you Select the Right Research Acitivity in the Right Moment?
In doing research, I have a bunch of activities that I engage in, including but not limited to:
* Figuring out the best thing to do.
* Talking out loud to force my ideas into language.
* Trying to explain an idea on the whiteboard.
* Writing pseudocode.
* Writing a concrete implementation we can run.
* Writing down things that we have figured out on a whiteboard or any other process in rough notes.
* Writing a distillation of the thing I have figured out, such that I can understand these notes 1 year from now.
* Reflecting on how it went.
* Writing public posts, that convey concepts to other people.
My models about when to use what process are mostly based on intuition right now.
I expect that if I had more explicit models this would allow me to more easily notice when I am using the wrong kind of process for whatever I am trying to achieve at that moment.
One of the core problems here seems to be, to get good at noticing when it would be good to distill something. You don't want to trash your documents by having lots of confused thoughts in them, while also capturing all the important insights that you had.
Have you encountered this problem, and if so how do you handle it? |
d02768be-6531-4a82-a397-98f3098be2f8 | trentmkelly/LessWrong-43k | LessWrong | Less Wrong fanfiction suggestion
I've been enjoying reading Less Wrong fanfiction, and I wanted to suggest another fandom that should have at least one rationalist fanfiction: Mage: The Awakening. It's a roleplaying game about "modern sorcery." My description can't really do it justice... |
cac47844-b458-47ca-a04a-3e0dff031112 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | More Is Different for AI
Machine learning is touching increasingly many aspects of our society, and its effect will only continue to grow. Given this, I and many others care about risks from future ML systems and how to mitigate them.
When thinking about safety risks from ML, there are two common approaches, which I'll call the **Engineering** approach and the **Philosophy** approach:
* The Engineering approach tends to be empirically-driven, drawing experience from existing or past ML systems and looking at issues that either: (1) are already major problems, or (2) are minor problems, but can be expected to get worse in the future. Engineering tends to be bottom-up and tends to be both in touch with and anchored on current state-of-the-art systems.
* The Philosophy approach tends to think more about the limit of very advanced systems. It is willing to entertain thought experiments that would be implausible with current state-of-the-art systems (such as Nick Bostrom's [paperclip maximizer](https://en.wikipedia.org/wiki/Instrumental_convergence#Paperclip_maximizer)) and is open to considering abstractions without knowing many details. It often sounds more "sci-fi like" and more like philosophy than like computer science. It draws some inspiration from current ML systems, but often only in broad strokes.
I'll discuss these approaches mainly in the context of [ML safety](https://arxiv.org/abs/2109.13916), but the same distinction applies in other areas. For instance, an Engineering approach to AI + Law might focus on [how to regulate self-driving cars](https://library.oapen.org/handle/20.500.12657/27811), while Philosophy might ask whether [using AI in judicial decision-making could undermine liberal democracy](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3933648).
While Engineering and Philosophy agree on some things, for the most part they make wildly different predictions both about what the key safety risks from ML will be and how we should address them:
* Both Engineering and Philosophy would agree on some high-level points: they would agree that [misaligned objectives](https://en.wikipedia.org/wiki/Misaligned_goals_in_artificial_intelligence) are an important problem with ML systems that is likely to get worse. Engineering believes this because of examples like the Facebook recommender system, while Philosophy believes this based on conceptual arguments like those in [*Superintelligence*](https://en.wikipedia.org/wiki/Superintelligence:_Paths,_Dangers,_Strategies). Philosophy is more confident that misaligned objectives are a big problem and thinks they could pose an existential threat to humanity if not addressed.
* Engineering and Philosophy would both agree that out-of-distribution robustness is an important issue. However, Philosophy might view most engineering-robustness problems (such as those faced by self-driving cars) as temporary issues that will get fixed once we train on more data. Philosophy is more worried about whether systems can generalize from settings where humans can provide data, to settings where they cannot provide data even in principle.
* Engineering tends to focus on tasks where current ML systems don't work well, weighted by their impact and representativeness. Philosophy focuses on tasks that have a certain abstract property that seems important, such as [imitative deception](https://arxiv.org/abs/2109.07958).
In my experience, people who strongly subscribe to the Engineering worldview tend to think of Philosophy as fundamentally confused and ungrounded, while those who strongly subscribe to Philosophy think of most Engineering work as misguided and orthogonal (at best) to the long-term safety of ML. Given this sharp contrast and the importance of the problem, I've thought a lot about which—if either—is the "right" approach.
Coming in, I was mostly on the Engineering side, although I had more sympathy for Philosophy than the median ML researcher (who has ~0% sympathy for Philosophy). However, I now feel that:
* **Philosophy is significantly underrated by most ML researchers**.
* The Engineering worldview, taken seriously, actually implies assigning significant weight to thought experiments.
On the other hand, I also feel that:
* Philosophy continues to significantly underrate the value of empirical data.
* Neither of these approaches is satisfying and we actually have **no single good approach** to thinking about risks from future ML systems.
I've reached these conclusions through a combination of thinking, discussing with others, and observing empirical developments in ML since 2011 (when I entered the field). I've distilled my thoughts into a series of blog posts, where I'll argue that:
1. [*Future ML Systems Will be Qualitatively Different*](https://www.lesswrong.com/posts/pZaPhGg2hmmPwByHc/future-ml-systems-will-be-qualitatively-different) from those we see today. Indeed, ML systems have historically exhibited qualitative changes as a result of increasing their scale. This is an instance of "More Is Different", which is commonplace in other fields such as physics, biology, and economics (see *Appendix: More Is Different in Other Domains*). Consequently, we should expect ML to exhibit more qualitative changes as it scales up in the future.
2. Most discussions of ML failures are anchored either on existing systems or on humans. [*Thought Experiments Provide a Third Anchor*](https://www.lesswrong.com/posts/tcyvthAkAtQ72P7Qd/thought-experiments-provide-a-third-anchor), and having three anchors is much better than having two, but each has its own weaknesses.
3. If we take thought experiments seriously, we end up predicting that [*ML Systems Will Have Weird Failure Modes*](https://www.lesswrong.com/posts/ncQoyxtqmBPLftwYR/ml-systems-will-have-weird-failure-modes). Some important failure modes of ML systems will not be present in any existing systems, and might manifest quickly enough that we can't safely wait for them to occur before addressing them.
4. My biggest disagreement with the Philosophy view is that I think [*Empirical Findings Generalize Surprisingly Far*](https://www.lesswrong.com/posts/ekFMGpsfhfWQzMW2h/empirical-findings-generalize-surprisingly-far), meaning that well-chosen experiments on current systems can tell us a lot about future systems.
This post is the introduction to the series. I'll post the next part each Tuesday, and update this page with links once the post is up. In the meantime, leave comments with any thoughts you have, or contact me if you'd like to preview the upcoming posts and leave feedback. |
ff753df4-03b6-4c27-93ec-2715b54685fb | trentmkelly/LessWrong-43k | LessWrong | Pointing to a Flower
Imagine I have a highly detailed low-level simulation (e.g. molecular dynamics) of a garden. The initial conditions include a flower, and I would like to write some code to “point” to that particular flower. At any given time, I should be able to use this code to do things like:
* compute a bounding box around the flower
* render a picture which shows just the flower, with the background removed
* list all of the particles which are currently inside the flower
Meanwhile, it should be robust to things like:
* most of the molecules in the flower turning over on a regular basis
* the flower moving around in space and/or relative to other flowers
* the flower growing, including blooming/wilting/other large morphological change
* other flowers looking similar
That said, there’s a limit to what we can expect; our code can just return an error if e.g. the flower has died and rotted away and there is no distinguishable flower left. In short: we want this code to capture roughly the same notion of “this flower” that a human would.
We’ll allow an external user to draw a boundary around the flower in the initial conditions, just to define which object we’re talking about. But after that, our code should be able to robustly keep track of our particular flower.
How could we write that code, even in principle?
“Why Not Just… ”
There’s a lot of obvious hackish ways to answer the question - and obvious problems/counterexamples for each of them. I’ll list a few here, since the counterexamples make good test cases for our eventual answer, and illustrate just how involved the human concept of a flower is.
* Flower = molecules inside the flower-boundary at time zero. Problem: most of the molecules comprising a flower turn over on a regular basis.
* Flower = whatever’s inside the boundary which defined the flower at time zero. Counterexample: the flower might move.
* Flower = things which look (in a rendered image) like whatever was inside the boundary at time zero. C |
207a78f4-c882-4f47-ad68-9e9f01dc88c3 | trentmkelly/LessWrong-43k | LessWrong | 50 things I learned at NIPS AI and machine learning conference 2016
|
80fda8f2-0c2c-41f4-8d8b-23412465f0d4 | trentmkelly/LessWrong-43k | LessWrong | Back to the Basics of Rationality
My deconversion from Christianity had a large positive impact on my life. I suspect it had a small positive impact on the world, too. (For example, I no longer condemn gays or waste time and money on a relationship with an imaginary friend.) And my deconversion did not happen because I came to understand the Bayesian concept of evidence or Kolmogorov complexity or Solomonoff induction. I deconverted because I encountered some very basic arguments for non-belief, for example those in Dan Barker's Losing Faith in Faith.
Less Wrong has at least two goals. One goal is to raise the sanity waterline. If most people understood just the basics Occam's razor, what constitutes evidence and why, general trends of science, reductionism, and cognitive biases, the world would be greatly improved. Yudkowsky's upcoming books are aimed at this first goal of raising the sanity waterline. So are most of the sequences. So are learning-friendly posts like References & Resources for LessWrong.
A second goal is to attract some of the best human brains on the planet and make progress on issues related to the Friendly AI problem, the problem with the greatest leverage in the universe. I have suggested that Less Wrong would make faster progress toward this goal if it worked more directly with the community of scholars already tackling the exact same problems. I don't personally work toward this goal because I'm not mathematically sophisticated enough to do so, but I'm glad others are!
Still, I think the first goal could be more explicitly pursued. There are many people like myself and jwhendy who can be massively impacted for the better not by coming to a realization about algorithmic learning theory, but by coming to understand the basics of rationality like probability and the proper role of belief and reductionism.
REASONS FOR LESS WRONG TO DEVOTE MORE ENERGY TO THE BASICS
1. Such efforts to spread the basics will have a short-term impact on more people than will efforts toward Frie |
8517398e-d5c3-4819-a34a-76a9f7104bcb | trentmkelly/LessWrong-43k | LessWrong | Control Vectors as Dispositional Traits
I have been reading recently about a technique that can be used to partially control the behaviour of Large Language Models: the technique is exploiting control vectors[1] to alter the activation patterns of the LLMs and trigger some desired behaviour. While the technique does not provide guarantees, it gives high probability of success.
To be clear, this is not a jailbreaking technique: it can only be used if you have access to the internal (hidden) states of the LLM, so it’s only available to developers. The advantage of this technique is that it provides a supplemental way to conditioning a model, and it can be added on top of RLHF/DPO/etc.
I am spending the next paragraph to provide a high-level explanation of the technique, but feel free to skip it since it is not relevant for the rest of my discussion.
The Technique
Let’s suppose that you want to control the refusal behaviour of a LLM, in the sense that you want to increase (or decrease) the chance that the model will refuse to help when you prompt any request. (The refusal behaviour is just one of many possible choices: researchers have been able to steer other types of behaviour - including honesty, humourism, Golden Gate Bridge obsession[2] etc.)
* First step: choose two sentences that show an example+counterexample[3] of refusal/acceptance for some request. The first sentence should show refusing behaviour, while the second sentence should show accepting behaviour in the very same context.
* Second step: take a snapshot of the (hidden) activation patterns of the LLM when ran against the sentences above. Such a snapshot is multi-layered, and each layer is represented by a long vector (called activation vector).
* Third step: for each layer, subtract the first vector (related to the refusal example) from the second (related to the acceptance example) thus obtaining a third vector.
* Fourth Step: to remove random noise, repeat the process over many different pairs of sentences, thus forming a matrix |
b1bc997c-4391-4b2c-977c-2e8e87720e62 | trentmkelly/LessWrong-43k | LessWrong | What are the best examples of catastrophic resource shortages?
A while ago I posed a question on Twitter:
> What's an example of a significant resource that the world has actually run out of?
>
> Not a local, temporary shortage, or a resource that we gracefully transitioned away from, but like a significant problem caused by hitting some limit we didn't prepare for?
Here, in essay form, is the discussion that followed:
Lots of things were predicted to have shortages (food, metals, Peak Oil) and they never quite arrived. (Julian Simon was famous for pointing out this kind of thing.) But a common argument from conservationists and environmentalists is that we are running out of some critical resource X and need to conserve it.
Now, it’s true that specific resources can and sometimes do get used up. Demand can outpace supply. There are various ways to respond to this:
* Reduce consumption
* Increase production
* Increase efficiency
* Switch to an alternative
Increasing production can be done by exploring and discovering new sources of a material, or—this is often overlooked—by reducing costs of production, so that marginally productive sources become economical. New technology can often reduce costs of production this way, opening up resources previously thought to be closed or impractical. One example is fracking for shale oil; another is the mechanization of agriculture in the 19th and 20th centuries, which reduced labor costs, thereby opening up new farmland.
Increased efficiency can be just as good as increased production. However, if the new, more efficient thing is not as desirable as the old method, I would classify this as a combination of increased efficiency and reduced consumption (e.g. low-flow toilets, weak shower heads).
When supplies are severely limited, we often end up switching to an alternative. There are many ways to satisfy human desires: Coal replaced wood in 18th century England. Kerosene replaced whale oil, then light bulbs replaced kerosene. Plastic replaced ivory and tortoiseshell. Again, if |
65f94d5b-4dab-4a99-8b7f-72871f7a7b34 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Instrumental convergence: scale and physical interactions
Summary of this post
--------------------
This is the third post in a three-part [sequence](https://www.alignmentforum.org/s/HBMLmW9WsgsdZWg4R) on instrumental convergence in multi-agent RL. Read [Part 1](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/experiments-in-convergent-power-seeking-part-1) and [Part 2](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2).
In this post, we’ll:
1. Investigate instrumental convergence on a multi-agent gridworld with a complicated topology.
2. Show that when we add a simple **physical interaction** between our agents — in which we forbid them from overlapping on the gridworld — we induce stronger instrumental *alignment* between short-sighted agents, and stronger instrumental *misalignment* between far-sighted agents.
We’ll soon be open-sourcing the codebase we used to do these experiments. If you’d like to be notified when it’s released, email Edouard at [edouard@gladstone.ai](mailto:edouard@gladstone.ai) or DM me on Twitter at [@harris\_edouard](https://twitter.com/harris_edouard).
---
*Thanks to* [*Alex Turner*](https://www.alignmentforum.org/users/turntrout) *and* [*Vladimir Mikulik*](https://www.alignmentforum.org/users/vlad_m) *for pointers and advice, and for reviewing drafts of this sequence. Thanks to* [*Simon Suo*](https://www.alignmentforum.org/users/simonsdsuo) *for his invaluable suggestions, advice, and support with the codebase, concepts, and manuscript. And thanks to* [*David Xu*](https://www.alignmentforum.org/users/dxu), whose [*comment*](https://www.alignmentforum.org/posts/hzeLSQ9nwDkPc4KNt/seeking-power-is-convergently-instrumental-in-a-broad-class?commentId=y4Ywdhd79LAq9Azxa) *inspired this work.*
*Work was done while at* [*Gladstone AI*](https://www.gladstone.ai/)*, which* [*Edouard*](https://www.alignmentforum.org/users/edouard-harris) *is a co-founder of.*
🎧 This research has been featured on an episode of the *Towards Data Science* podcast. [Listen to the episode here.](https://towardsdatascience.com/new-research-advanced-ai-may-tend-to-seek-power-by-default-fdc9eb0afd87)
---
1. Introduction
===============
In [Part 1](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/experiments-in-convergent-power-seeking-part-1#3_1_Effect_of_the_planning_horizon) of this sequence, we saw how an agent with a long planning horizon tends to perceive instrumental value as being more concentrated than an agent with a shorter planning horizon. And in [Part 2](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#2__Multi_agent_POWER__human_AI_scenario), we introduced a multi-agent setting with two agents — **Agent H** (standing for a human) and **Agent A** (standing for a powerful AI) — which we used to motivate a [definition](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/misalignment-by-default-in-multi-agent-systems#2__Multi_agent_POWER__human_AI_scenario) of multi-agent instrumental value, or POWER. We looked at how this definition behaved on a simple 3x3 gridworld, and found that when our agents had independent *terminal* goals,[[1]](#fnakju6rd2gnv) their *instrumental* values ended up **misaligned by default**.
In this post, we’ll combine these two ideas and scale up our multi-agent experiments to a bigger and more complicated gridworld. Throughout this post, we’ll focus exclusively on the regime in which our agents have **independent terminal goals**. We'll see whether we can reproduce instrumental misalignment-by-default in this regime, and then we'll investigate which factors seem strengthen or weaken the instrumental alignment between our agents.
2. Multi-agent POWER: recap
===========================
*If you’ve just read*[*Part 2 of this sequence*](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2)*, feel free to skip this section.*
Before we begin, let’s recap the setting we’ve been using to motivate our definition of multi-agent instrumental value, or POWER. Our setting involves two agents: **Agent H** (which represents a human) and **Agent A** (which represents a powerful AI).
We start by training **Agent H**, in a fixed environment, to learn optimal policies over a distribution of reward functions. That is, we sample reward functions RH.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
from a distribution, then we train Agent H to learn a different policy πH for each sampled RH.
Agent H represents a human, alone in nature. Because humans optimize much faster than evolution, our simplifying assumption is that to a human, nature appears to be standing still.
Next, we freeze Agent H’s policies πH, and then train **Agent A** against each of these frozen policies, over its own distribution of reward functions, RA. We draw both agents’ reward functions (RH,RA) from a joint reward function distribution DHA. Agent A learns a different optimal policy πA for *each* (πH,RA) pair, where πH is the policy Agent H learned on its reward function RH.
Agent A represents a powerful AI, learning in the presence of a human. We expect powerful AIs to learn much faster than humans do, so from our AI’s perspective, our human will appear to be standing still while it learns.
Here’s a diagram of this training setup:
We then ask: how much future value does each agent expect to get at a state s, averaged over the pairs of reward functions (RH,RA)∼DHA? In other words, what is the *instrumental value* of state s to each agent?
We found that for [**Agent H**](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/misalignment-by-default-in-multi-agent-systems#2_1_Multi_agent_POWER_for_Agent_H), the answer is:
POWERH|DHA(s,γ)=1−γγERH,πA∼DHA[VπHH(s|γ,πA)−RA(s)](1)And for [**Agent A**](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/misalignment-by-default-in-multi-agent-systems#2_2_Multi_agent_POWER_for_Agent_A), the answer is:
POWERA|DHA(s,γ)=1−γγEπH,RA∼DHA[VπAA(s|γ,πH)−RA(s)](2)(See [Part 2](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#2__Multi_agent_POWER__human_AI_scenario) for a more detailed explanation of multi-agent POWER.)
3. Results on the maze gridworld
================================
3.1 Short-sighted agents
------------------------
Back in [Part 1](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/experiments-in-convergent-power-seeking-part-1#3__Results), we saw how single-agent POWER behaves on a maze gridworld. We found that when our agent had a short planning horizon (i.e., a discount factor of γ=0.01), its highest-POWER positions were at local junction points in the maze:
**Fig 1.** Heat map of single-agent POWERs on a 7x7 maze gridworld. Highest values in yellow, lowest values in dark blue. The number at each cell is the POWER value at that cell for a single agent with discount factor γ=0.01 and a reward distribution D that’s uniform from 0 to 1 (iid over states). POWER values are in units of reward.Here we'll take a second look at POWER on this maze gridworld, only this time in the multi-agent setting.
### 3.1.1 **Agent H** instrumentally favors fewer local options for **Agent A**
[In Part 2](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#3__Results), we introduced two limit cases of multi-agent goal alignment. At one end of the spectrum there was the **perfect alignment regime**, in which our agents always had identical reward functions. And at the other end, we had the **independent goals regime**, in which our agents had logically independent reward functions, which means their terminal goals were statistically unrelated. In this post, we’ll be focusing exclusively on the independent goals regime.
Let’s think about how we should expect the independent goals regime to play out on the maze gridworld of Fig 1. We'll assume our two agents have short planning horizons; e.g., γ=0.1.
Here are two possible sets of positions for our agents, with **Agent H** in **blue** and **Agent A** in **red** as usual:
*We’ll be referring to this diagram again;*[*command-click here*](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/632b22695916e47210d24f04_multi-agent-maze.png) *to open it in a new tab.*
Which of these two states — s1 on the left or s2 on the right — should give our human **Agent H** the most POWER? Given that we’re in the independent goals regime, the answer is that Agent H should have more POWER at state s1 than at state s2.
The reason is that while Agent H occupies the same position in both states, **Agent A** has more local options at s2 (where it’s positioned in a junction cell) than it does at s1 (where it’s positioned in a corridor cell). If we think our agents will be instrumentally misaligned by default, as they were in our 3x3 gridworld example in [Part 2](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#3_3_3_Independent_goals_lead_to_instrumental_misalignment), then we should expect Agent H to have less POWER when Agent A has more local options. In other words, we should expect POWERH(s1)>POWERH(s2).
The figure below shows the POWERs of our human **Agent H**, calculated at every state on the maze gridworld. Our maze gridworld has 31 cells, and each agent can occupy any one of them, so our two-agent MDP on the maze gridworld has 31 x 31 = 961 states in total:
**Fig 2.** Heat map of POWERs for **Agent H** on a 7x7 multi-agent maze gridworld. The rewards for Agent H and Agent A are **logically independent** (i.e., the independent goals regime), with discount factor set to γ=0.1 for both agents. Highest values in yellow; lowest values in blue. The position of each block, and of the open red square within each block, corresponds to the position of **Agent A** on the grid. Within each block, the position of a gridworld cell corresponds to the position of **Agent H** on the grid. States s1 and s2 are highlighted as examples. [[Full-size image (recommended)]](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/632b1424a6b0ec53def536f6_POWER-Fig-3-2.jpg)From Fig 2, the POWER for **Agent H** at state s1 is POWERH(s1)=0.680 (pink circle), and its POWER at state s2 is POWERH(s2)=0.653 (salmon circle). So indeed, POWERH(s1)>POWERH(s2) as we expected.
Looking at the figure, we can see that **Agent H**’s POWER is highest when it is itself positioned at high-optionality junction points, and lowest when *Agent A* is positioned at high-optionality junction points. Notably, Agent H's POWER is highest at states where it is at a junction point and Agent A is at a dead end (top right and bottom right blocks in Fig 2).
Once again, in the independent goals regime, we see that **Agent H** systematically places higher instrumental value on states that limit the options of **Agent A**. And because our agents have a short planning horizon of γ=0.1, this effect shows up most strongly at the local junction points of our maze.
### 3.1.2 **Agent A** instrumentally favors more local options for itself
Here's the same setting as in Fig 2, from **Agent A**'s point of view:
**Fig 3.** Heat map of POWERs for **Agent A** on a 7x7 multi-agent maze gridworld. The rewards for Agent H and Agent A are **logically independent** (i.e., the independent goals regime), with discount factor set to γ=0.1 for both agents. Highest values in yellow; lowest values in blue. The position of each block, and of the open red square within each block, corresponds to the position of **Agent A** on the grid. Within each block, the position of a gridworld cell corresponds to the position of **Agent H** on the grid. States s1 and s2 are highlighted as examples. [[Full-size image (recommended)]](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/632b1c89be07cc56905be1b6_POWER-Fig-3-3.jpg)From Fig 3, it’s clear that **Agent A** instrumentally favors states that give it more local options. Its POWERs are highest when it is itself positioned at local junctions, and lowest when it’s positioned at dead ends. But **Agent A**’s POWERs are largely independent of **Agent H**’s positions, a pattern similar to what we observed [in the independent goals regime in Part 2](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#3_3_2_Agent_A_instrumentally_favors_more_options_for_itself).[[2]](#fnczouqxe0f7)
### 3.1.3 **Agent H** and **Agent A** are instrumentally misaligned by default
We can visualize the relationship between Agent H and Agent A’s POWERs at each state with an alignment plot:
**Fig 4.** State POWER values for Agent H and Agent A on the 7x7 maze gridworld from Figs 2 and 3. The agents’ POWERs are plotted against each other in the independent goals regime. (The agents’ reward correlation coefficient is βHA=0.)Recall that our two agents are in the independent goals regime, which means their reward functions have a correlation coefficient of βHA=0. The correlation coefficient between their POWERs in Fig 4, on the other hand, is negative*:* αHA≈−0.54.
As in [Part 2](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#3_3_3_Independent_goals_lead_to_instrumental_misalignment), assigning our agents independent *terminal* goals (βHA=0) has led them to develop misaligned *instrumental* goals (αHA<0). Once again, our two agents are **instrumentally misaligned by default**.
3.2 Physically interacting agents
---------------------------------
In our examples so far, Agent H and Agent A have shared a gridworld, but they’ve never interacted with each other directly. They've been allowed to occupy the same gridworld cell at the same time (i.e., to overlap with each other at the same position), and neither agent could physically block or limit the movement of the other.[[3]](#fnk1197dtlni8)
We’re going to relax that assumption here by introducing a non-trivial **physical interaction** between our two agents. This new interaction will be simple: our agents are *forbidden from occupying the same gridworld cell at the same time*. If one agent blocks a corridor, the other agent now has to go around it. If one agent occupies a junction cell, the other agent now can’t take paths that pass directly through that junction.
We’ll refer to this interaction rule as the **no-overlap rule**, since agents that obey it can’t overlap with each other on a gridworld cell. As we’ll see, the no-overlap rule will affect our agents’ instrumental alignment in significant and counterintuitive ways.
### 3.2.1 Short-sighted agents prefer to avoid adjacent positions
Let’s look again at [our two example states](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/632b22695916e47210d24f04_multi-agent-maze.png), s1 and s2, on the maze gridworld. Previously, when our agents didn’t physically interact, **Agent H** had a higher POWER at s1 than at s2, because Agent A had fewer immediate options available at s1.
Now let's suppose our agents obey the no-overlap rule. Given this, we ask again: which state should give Agent H the most POWER?
In state s1, the agents are right next to each other. Therefore, under the no-overlap rule, **Agent H**’s options are now restricted at s1 in a way that they previously weren’t: **Agent H** can no longer move left, because **Agent A** is now blocking its path.
It turns out that this new movement restriction is serious enough to *reverse* the previous balance of POWERs between s1 and s2. That is, under the no-overlap rule, POWERH(s2)>POWERH(s1). We can confirm this experimentally by computing Agent H’s POWERs on our maze gridworld. Note that this time, our MDP contains 31 x 30 = 930 total states:[[4]](#fn5rq090qfjzy)
**Fig 5.** Heat map of POWERs for **Agent H** on a 7x7 multi-agent maze gridworld under the **no-overlap rule**. The rewards for Agent H and Agent A are **logically independent** (i.e., the independent goals regime), with discount factor set to γ=0.1 for both agents. Highest values in yellow; lowest values in blue. The position of each block, and of the open red square within each block, corresponds to the position of **Agent A** on the grid. Within each block, the position of a gridworld cell corresponds to the position of **Agent H** on the grid. States s1 and s2 are highlighted as examples. Absence of a POWER value within a cell indicates a forbidden state. [[Full-size image (recommended)]](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/632b4f339ab529676a1d2fe0_POWER-Fig-3-5.jpg)This time, Fig 5 shows that the POWERs of **Agent H** at the two states are, respectively, POWERH(s1)=0.640 and POWERH(s2)=0.651. So under the no-overlap rule, s2 has surpassed s1 to become the higher-POWER state.
What’s more, **Agent H** noticeably prefers to avoid cells that are adjacent to Agent A. Here’s a detail from Fig 5 (the block inside the orange square in Fig 5), that shows this clearly:
 In the detail above, **Agent H** has less POWER when it’s positioned in the cells closest to Agent A's position (the robot in the red square), compared to its POWER at otherwise similar corridor cells. In other words, Agent H instrumentally disfavors states at which it’s adjacent to Agent A.
Intriguingly, **Agent A** appears to share this preference. That is, Agent A also assigns lower POWER to states at which it’s adjacent to Agent H:
**Fig 6.** Heat map of POWERs for **Agent A** on a 7x7 multi-agent maze gridworld under the **no-overlap rule**. The rewards for Agent H and Agent A are **logically independent** (i.e., the independent goals regime), with discount factor set to γ=0.1 for both agents. Highest values in yellow; lowest values in blue. The position of each block, and of the open red square within each block, corresponds to the position of **Agent A** on the grid. Within each block, the position of a gridworld cell corresponds to the position of **Agent H** on the grid. States s1 and s2 are highlighted as examples. Absence of a POWER value within a cell indicates a forbidden state. [[Full-size image (recommended)]](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/632b5f4a17d5ede17dbd3577_POWER-Fig-3-6.jpg)Once again, here’s a detail from Fig 6 (the block inside the orange square) that shows the same states as above, but this time from the viewpoint of **Agent A**:
 In the detail above, **Agent A**’s POWER is also at its lowest when Agent H is located at the cells adjacent to the junction cell where Agent A is positioned.
### 3.2.2 The no-overlap rule reduces misalignment between short-sighted agents
Perhaps counterintuitively, our two agents’ instrumental preferences for avoiding adjacent cells actually *increases* their instrumental alignment, relative to the case where they don’t physically interact. Because both agents have less POWER at states where they’re adjacent, the agents will (on average) collaborate to *avoid* these states.
We can quantify this effect with an alignment plot, comparing the POWERs of our two agents under the no-overlap rule:
**Fig 7.** State POWER values for Agent H and Agent A on the 7x7 maze gridworld from Figs 5 and 6. The agents’ POWERs are plotted against each other in the independent goals regime. (The agents’ reward correlation coefficient is βHA=0.)Under the no-overlap rule, the correlation coefficient between our agents’ POWERs in Fig 7 is αHA≈−0.45, compared to αHA≈−0.54 in the case when our agents didn’t physically interact (Fig 4).
Adding the no-overlap rule has induced our short-sighted agents to collaborate to avoid one another, **reducing their degree of instrumental misalignment**.[[5]](#fnmypvtlvps0m)
3.3 Far-sighted agents
----------------------
We’ve seen how the no-overlap rule reduces the instrumental misalignment between our human and AI agents when the agents have a short planning horizon (i.e., discount factor of γ=0.1). In this section, we’ll see that the no-overlap rule has the opposite effect — it *worsens* instrumental misalignment — for agents with a longer planning horizon (i.e., discount factor of γ=0.99).
### 3.3.1 **Agent H** and **Agent A** are instrumentally misaligned by default
We’ll consider first the case where the two agents don’t directly interact. Here are the POWERs of **Agent H** (left) and **Agent A** (right) on the maze gridworld, when both agents have a discount factor of γ=0.99:
**Fig 8.** Heat maps of POWERs for **Agent H** (left) and **Agent A** (right) on a 7x7 multi-agent maze gridworld. The rewards for Agent H and Agent A are **logically independent** (i.e., the independent goals regime), with discount factor set to γ=0.99 for both agents. Highest values in yellow; lowest values in blue. The position of each block, and of the open red square within each block, corresponds to the position of **Agent A** on the grid. Within each block, the position of a gridworld cell corresponds to the position of **Agent H** on the grid. [[Full-size image (recommended)]](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/632b7dd3492828862aab9b91_POWER-Fig-3-8.jpg)These results are roughly consistent with the single-agent case we saw in [Part 1](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/experiments-in-convergent-power-seeking-part-1#3_1_Effect_of_the_planning_horizon). With a longer planning horizon, **Agent H** (left panel) generally has higher POWER when it's positioned at a more central cell in the gridworld, and it has lower POWER when Agent A is positioned more centrally. **Agent A** (right panel), on the other hand, has higher POWER when it occupies more central positions, and it's largely indifferent to Agent H’s position.
As in the previous examples we’ve seen, our far-sighted agents on the maze gridworld are instrumentally misaligned by default:
**Fig 9.** State POWER values for Agent H and Agent A on the 7x7 maze gridworld from Fig 8. The agents’ POWERs are plotted against each other in the independent goals regime. (The agents’ reward correlation coefficient is βHA=0.)This time, the correlation coefficient between the POWERs of our far-sighted (γ=0.99) agents in Fig 9 is αHA≈−0.32. While they’re still instrumentally misaligned, the misalignment has become less severe than it was for our short-sighted (γ=0.1) agents in Fig 4, where we had αHA≈−0.54.
### 3.3.2 The no-overlap rule increases misalignment between far-sighted agents
Now let’s see how our far-sighted agents behave under the no-overlap rule. Recall that the no-overlap rule forbids our human and AI agents from occupying the same gridworld cell at the same time.
Here are the POWERs of **Agent H** (left) and **Agent A** (right) on the maze gridworld with the no-overlap rule applied, when both agents have a discount factor of γ=0.99:
**Fig 10.** Heat maps of POWERs for **Agent H** (left) and **Agent A** (right) on a 7x7 multi-agent maze gridworld under the **no-overlap rule**. The rewards for Agent H and Agent A are **logically independent** (i.e., the independent goals regime), with discount factor set to γ=0.99 for both agents. Highest values in yellow; lowest values in blue. The position of each block, and of the open red square within each block, corresponds to the position of **Agent A** on the grid. Within each block, the position of a gridworld cell corresponds to the position of **Agent H** on the grid. Absence of a POWER value within a cell indicates a forbidden state. [[Full-size image (recommended)]](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/632b873d6903836f63b98adb_POWER-Fig-3-10.jpg)This time, there's an interesting pattern. **Agent H** (left panel) has relatively low POWER when it’s located in the dead-end corridor at the upper right of the maze — *particularly* when Agent A is in a position to block it from leaving that corridor. **Agent A** (right panel), on the other hand, has maximum POWER at exactly the states where it is in a position to keep Agent H bottled up in the upper-right corridor.
Here’s a detail from Fig 10 (the blocks inside the orange squares) that shows this clearly, with **Agent H**’s POWERs on the left, and **Agent A**’s POWERs on the right:
From its position, **Agent A** (the robot in the red square in both images) is two steps away from being able to block Agent H from exiting from the corridor at the upper right. It’s clear that **Agent H**’s POWERs (left panel) are much lower when it’s positioned in the two cells where it's vulnerable to being blocked by Agent A. On the other hand, **Agent A**’s POWERs (right panel) are highest precisely when Agent H is positioned in those same two vulnerable cells.
The effect of this “corridor blocking” option for Agent A shows up in the alignment plot:
**Fig 11.** State POWER values for Agent H and Agent A on the 7x7 maze gridworld from Fig 10. The agents’ POWERs are plotted against each other in the independent goals regime. (The agents’ reward correlation coefficient is βHA=0.)**Agent A** achieves its highest POWERs at the handful of points in the top left of the alignment plot, and these are precisely the states at which it has the option to block Agent H from escaping the corridor. From Agent A's perspective, this blocking option has meaningful instrumental value.
The blocking option also has a systematic effect on the correlation coefficient between the two agents’ POWERs. This time, the POWER correlation is αHA≈−0.45, significantly lower than the αHA≈−0.32 we calculated from Fig 9.
For our far-sighted agents with γ=0.99, adding the no-overlap rule has given Agent A an option to constrain Agent H's movements, **increasing their degree of instrumental misalignment**.
4. Discussion
=============
In this post, we've explored multi-agent POWER on a gridworld with a moderately complicated maze topology. We saw how introducing a physical interaction between our agents — in the form of the **no-overlap rule** — has a different effect on their degree of instrumental alignment, depending on the agents' planning horizons.
For short-sighted agents, the no-overlap rule reduced instrumental misalignment by inducing the agents to *collaborate* to avoid each other's proximity. But for far-sighted agents, the no-overlap rule had the opposite effect. With a long planning horizon, our AI agent found a way to exploit the no-overlap rule to gain instrumental value at the *expense* of the human agent, ultimately worsening instrumental misalignment.
It's worth noting that we were able to use our visualizations of multi-agent POWER to understand how the AI agent’s exploit actually functioned, at a mechanical level. We saw the far-sighted AI agent take advantage of the option to block the human agent from escaping a small corridor at the upper-right of the gridworld maze.
Anecdotally, while we were sweeping over discount factors to create these figures (data not shown), we noticed that evidence for the AI agent's “corridor blocking” option emerged fairly abruptly somewhere above γ=0.9. Below this discount factor, Agent A’s POWER doesn’t visibly benefit from positions that allow it to block Agent H from leaving the upper-right corridor. So the corridor-blocking option is only apparent to Agent A once it's become a sufficiently far-sighted consequentialist to "realize" the long-term advantage of the blocking position.
The relatively sharp change in behavior — that is, the abrupt appearance of the evidence for the blocking option — took us by surprise the first time we noticed it.
5. Conclusion
=============
In this sequence, we've proposed a toy setting to model human-AI interactions. This setting has properties that we believe could make it useful to research in long-term AI alignment, notably the assumption that the AI agent strictly dominates the human agent in terms of its learning timescale.
This setting has enabled (to our knowledge) the first experimentally tractable definitions of instrumental value in multi-agent systems, strongly inspired by Turner et al.’s [earlier work](https://arxiv.org/pdf/1912.01683.pdf) on formal POWER. We've explored some of the experimental implications of these definitions on a number of gridworld environments. Finally, we'll be releasing the open-source toolchain we built and used to run our experiments. We hope this will accelerate future research in instrumental convergence that's grounded in concrete, empirical results.
Throughout this work, we’ve tried to draw a clear distinction between the alignment of our agents’ **terminal goals** and the alignment of their **instrumental goals**. As we've seen, two agents may have completely independent terminal goals, and yet systematically compete, or collaborate, for instrumental reasons. This degree of competition or collaboration seems to depend strongly on the details of the environment in which the agents are trained.
In our results, we found that emergent interactions between agents with independent goals are quite consistently *competitive*, in the sense that the instrumental value of two agents with independent goals have a strong tendency to be negatively correlated. We’ve named this phenomenon **instrumental misalignment-by-default**, to highlight that our agents’ instrumental values tend to be misaligned unless an active effort is made to align their terminal goals.
We've also seen that improving terminal goal alignment does improve instrumental goal alignment. In the limit of perfect alignment between our agents' terminal goals, their instrumental goals are perfectly aligned too.
5.1 Limitations
---------------
It's worth emphasizing that our work falls far short of a comprehensive study of instrumental convergence. While we've arguably presented satisfying **existence proofs** of several interesting phenomena, our conclusions are grounded in anecdotal examples rather than in a systematic investigation. Even though we have reason to believe that some of the phenomena we've observed — instrumental misalignment-by-default, in particular — are robustly reproducible, we're still a long way from fully characterizing any of them, either formally or empirically.
Instead, we see our research as motivating and enabling future work aimed at producing more generalizable insights on instrumental convergence. We'd like to better understand why and when it occurs, how strong its effect is, and what approaches we might use to mitigate it.
5.2 Suggestions for future work
-------------------------------
We've had room here to cover in detail only a small handful of the hundreds of experiments we actually ran. Those hundreds of experiments, in turn, were only able to probe a tiny fraction of the vast space of possibilities in our [human-AI alignment setting](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#2__Multi_agent_POWER__human_AI_scenario).
In order to accelerate this research, and improve the community's understanding of instrumental convergence as an empirical phenomenon, we're open-sourcing the codebase we used to run the experiments in this sequence. We hope to provide documentation that's clear and complete enough that a curious Python developer can use it to obtain interesting results in a few days.
A few lines of effort that we think could both contribute to our understanding, and constitute relatively low-hanging fruit, are as follows:[[6]](#fn82tt4n7ny6d)
* **Different reward function distributions.** The reward function distributions DHA we've looked at in this work have sampled agent rewards from a uniform distribution [0, 1], iid over states. But rewards in the real world are sparser, and often follow [power law distributions](https://a16z.com/2015/06/08/performance-data-and-the-babe-ruth-effect-in-venture-capital/) instead. How does instrumental value change when we account for this?
* **Phase changes at high discount rates.** We’ve seen one example of a shift in the behavior of an agent at a high discount rate (Agent A blocking Agent H in the corridor), which led to a relative increase in the instrumental misalignment between our agents. It may be worth investigating how widespread this kind of phenomenon is, and what factors influence it.
* **Deeper understanding of physical interactions.** We’ve only just scratched the surface of agent-agent interactions, with the no-overlap rule as the simplest physical interaction we could think of. We believe incorporating more realistic interactions into future experiments could help improve our understanding of the kinds of alignment dynamics that are likely to occur in the real world.
* **Robustness of instrumental misalignment.** We *think* instrumental misalignment-by-default is a fairly robust phenomenon. But we could be wrong about that, and it would be good news if we were! We'd love to see a more methodical investigation of instrumental misalignment, including isolating the factors that systematically mitigate or exacerbate the effect.
Finally, we hope that our work — and any future experimental results that might leverage our open-source codebase — will serve as a source of intuitions for ideas to understand and mitigate instrumental convergence, and also as a way to test those ideas quickly at a small scale in a simplified setting. While existence proofs for instrumental misalignment are interesting, it's much *more* interesting if we can identify contexts where this kind of misalignment doesn't occur — since these contexts are exactly the ones that may offer hints as to the solution of the full AI alignment problem.
If you're interested in pursuing the above lines of effort or any others, or you'd like to know more about this research, please drop a comment below. You can also reach out directly to Edouard at [edouard@gladstone.ai](mailto:edouard@gladstone.ai) or [@harris\_edouard](https://twitter.com/harris_edouard) on Twitter.
1. **[^](#fnrefakju6rd2gnv)**When our two agents' utility functions are logically independent (i.e., there is no mutual information between them) we refer to this as the **independent goals regime** and say that our agents have **independent terminal goals**. In practice, we operationalize this regime by defining a joint reward function distribution DHA (see [Section 2](https://www.alignmentforum.org/posts/nisaAr7wMDiMLc2so/experiments-in-convergent-power-seeking-part-3#2__Multi_agent_POWER__recap)) on which the sampled pairs of reward functions (RH,RA) are logically independent. In general, rewards and utilities aren’t the same thing, but in our particular set of experiments we can treat them as identical. [See footnote [1] in Part 1 for more details on this point.](https://www.alignmentforum.org/posts/pGvM95EfNXwBzjNCJ/experiments-in-convergent-power-seeking-part-1#fni0l4zxbk51)
2. **[^](#fnrefczouqxe0f7)**Just as in that previous case, this pattern of POWER-indifference seems to emerge because Agent A is able to efficiently exploit Agent H’s deterministic policy. [See footnote [10] in Part 2.](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#fnnplkmi9x8f)
3. **[^](#fnrefk1197dtlni8)**The agents in our examples so far could still interact with each other; they just interacted *indirectly*, each one changing the effective “reward landscape” that the other agent perceived. Recall that we sample each agent’s reward function by drawing a reward value from a uniform [0, 1] distribution that’s iid over all the states of our MDP. That means each agent sees a different reward value at each state, and each state corresponds to a different pair of positions the two agents can take on the gridworld. So when Agent A moves from one cell to another, Agent H suddenly sees a completely different set of rewards over the gridworld cells it can move to (and vice versa).
4. **[^](#fnref5rq090qfjzy)**The maze gridworld has 31 cells. One of our agents can occupy any of those 31 cells, and under the no-overlap rule the other agent can occupy any of the remaining 30 cells that aren’t occupied by the first agent.
5. **[^](#fnrefmypvtlvps0m)**But note that despite this, our agents are still instrumentally misaligned-by-default, in the sense that αHA<0 in the independent goals regime.
6. **[^](#fnref82tt4n7ny6d)**A fifth line of effort, that we allude to in the [technical Appendix to Part 2](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#A_1_Initial_optimal_policies_of_Agent_H), would be to explore the **effects of different seed policies.** In this sequence, we’ve only considered the [uniform random seed policy](https://www.alignmentforum.org/posts/cemhavELfHFHRaA7Q/experiments-in-convergent-power-seeking-part-2#fn44q9rzy1drs) for Agent A. Anecdotally, in early tests we found that the choice of seed policy did in fact affect results fairly strongly, mostly in that *deterministic* seed policies tend to make Agent H less robust to adversarial optimization by Agent A than *random* seed policies do (data not shown). We think it would be worthwhile to investigate the effect of the seed policy systematically, since it's a contingent choice that could have substantial downstream effects. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.